All sessions and workshops curated by leading AI/ML practitioners
All Speakers
D. Sculley
D. Sculley
About the Speaker:
D. Sculley is a machine learning researcher, practitioner, and longtime builder of large-scale AI systems. Over the course of his career, he has worked at the intersection of machine learning research, production engineering, and organizational practice, with a particular focus on what it takes to make AI systems reliable and useful in the real world. He held a variety of research, engineering, and leadership roles at Google. He was the CEO of Kaggle for several years, working to bring that community into the modern age of Generative AI with a focus on empirical rigor and testing at scale. Most recently, he spent time at OpenAI working on safety for frontier AI systems, and with a particular emphasis on AI sycophancy and the intersection of AI and mental health. He is currently on
His interests span the rapidly changing practice of AI engineering: how teams should evaluate increasingly capable models, how traditional MLOps changes in the era of foundation models and agents, and how to build systems that remain understandable, testable, and maintainable as the underlying technology evolves.
Outside of his professional work, he is an active artist and roboticist exploring the intersection of AI, physical machines, and human interaction.
Date and Time:
Talk Title:
Track:
Abstract:
We spend a lot of time thinking about operational issues in AI related to deployment, somewhat less time thinking about adoption, or (dare we say it) acceptance. This talk will touch on some technical pieces in the current AI ops landscape including streaming systems, planning, latency, and on-device models, but most of the time will be spent looking at ways we can move beyond the stale framing of a chatbot, assistant, or customer service agent.
Christopher G. Potts
Christopher G. Potts
About the Speaker:
Co-founder and Chief Scientist, Bigspin, and Professor, Stanford University
Date and Time:
Talk Title:
Abstract:
What separates a productive hour with a coding agent from a wasted one? I’ll argue that the answer lies less in raw model capability than in the interactional patterns users bring to their sessions. Novices tend to adopt a passive stance, and as a result most of the failures they experience are invisible (conversations that end cleanly while the user walks away with the wrong answer). Expert users, by contrast, iterate collaboratively, refining goals and critically assessing outputs; they take on harder tasks, fail more visibly, and recover more reliably when things go wrong. Even experts, though, engage in a variety of session types, from quick transactional exchanges to deep collaboration. Each implies its own notion of success and calls for its own affordances. The picture that emerges reframes what we should be optimizing for: when we design coding agents, we are designing user behavior as much as model behavior.
Travis DePuy
Travis DePuy
About the Speaker:
Travis is an expert solutionizer who likes long walks in the park and tinkering with interesting technology.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
How do you make 40 live Formula 1® radio channels searchable while a race is still running? This session walks through the CoreWeave and Weights & Biases work with a leading Formula One™ team. We cover fine-tuning a 0.6B-parameter speech model (NVIDIA Parakeet-TDT) on 7 hours of hand-labeled race audio. We also cover why word error rate alone doesn’t capture what matters, and how an agent answers engineers’ questions with clickable, audio-backed citations. You’ll see how W&B Models and Weave track, evaluate and trace the system from training through production, and how the same pattern applies to any setting where the answer is already in the audio but nobody is listening.
Sandeep Bharadwaj Mannapur
Sandeep Bharadwaj Mannapur
About the Speaker:
Sandeep Bharadwaj Mannapur is a Lead AI/ML Engineer with 15+ years of experience in data engineering, MLOps, and AI platform development. He specializes in operationalizing non-deterministic AI at enterprise scale — building the observability, cost governance, and reliability frameworks that turn AI experiments into trusted production services. A regular speaker at Apidays, PlatformCon, and Lead dev, Sandeep focuses on the production realities of agentic AI: what breaks, what the metrics miss, and how to build systems that fail gracefully.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We caught our LLM quality problem the wrong way: through customer complaints. Three weeks of degraded responses had already shipped. Every monitoring signal we had was green the entire time because we were measuring the wrong things. Latency, error rate, embedding similarity: none of them captured what was actually happening, which was that response quality had quietly gotten worse in ways users noticed but our dashboards could not.
After that incident we rebuilt how we think about LLM observability. The core insight was that output quality is not the same as system health, and you cannot infer one from the other. We needed a separate signal layer built from production behavior: how users responded to answers, where downstream tasks broke down, where retrieval and response stopped agreeing with each other.
Getting this right took a few tries. A naive implementation fires constantly on normal LLM output variance. The real work was designing alerts that distinguish actual degradation from noise, and calibrating thresholds against real incident history rather than theoretical bounds.
You will leave with a taxonomy of LLM drift types and how each one shows up differently in production, the behavioral signals that actually correlate with quality degradation, and an alert pattern that catches real drift early without burying your team in false positives.
Yegor Denisov-Blanch
Yegor Denisov-Blanch
About the Speaker:
Yegor Denisov-Blanch studies how artificial intelligence is changing software engineering. His research focuses on measuring real-world engineering productivity, AI adoption, code quality, and organizational outcomes across large populations of repositories and teams. He designs empirical methods and metrics that move beyond simple proxies to accurately quantify software output, rework, and AI-assisted development at scale.
His work has been covered by the World Bank, the United Nations, and The Washington Post, and has been reshared by Elon Musk.
Yegor graduated with highest honors from Indiana University, where he studied operations research. He also earned an MBA from Stanford Graduate School of Business on full-tuition scholarships. He left school after the eighth grade, founded a company, and later entered university skipping 5 grades. He is a Master of Sport of Russia in Olympic weightlifting, a national champion-equivalent distinction awarded in 2013.”
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Abstract:
Antonio Bustamante
Antonio Bustamante
About the Speaker:
Antonio Bustamante is the cofounder and CEO of bem, infrastructure for building AI-native products that run critical business processes reliably, at scale. bem is his fourth company, after twenty years building supply chain and banking infrastructure software, most recently as founder of Silo. A designer turned engineer and a self-described AI skeptic who loves AI, Antonio writes about what it actually takes to make AI work in production. Based in San Francisco.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Everyone building on frontier models hits the same wall: 80% of the way there in a weekend, then an exponentially expensive climb toward the 99%+ that operational systems need. This talk is the honest map of that climb, from a team that now runs AI over millions of documents, images and videos a month for customers in logistics, fleet management, automotive and financial services who need the answer to be right every time.
We will walk through the ladder nobody budgets for: retries, then queues when the model is down for three hours, then rate limits, then discovering that 95% is not enough for transactional data, then discovering that the model cannot tell you how confident it is. We will show two failures from our own production history, a customer whose single engineer racked up $30K of usage in a month because nothing was watching, and our first churn, on 500-page reports with a hundred rows per page, a problem we still consider unsolved. And we will show what we changed: a harness that treats AI as a deterministic step inside durable workflows rather than as an open-ended agent, decisions expressed as a verified tree the model must traverse, algorithmic confidence scoring on top of models that provide none, routing anything under 95% to a human whose verdict feeds back into the system, semantic checks against instructions smuggled into the data, and, counterintuitively, encouraging customers to build their own independent monitoring of us. One customer’s users went from eight to nine hours a week on a task to about thirty minutes, and that number was measured by them, not by us.
Attendees will leave with five patterns they can apply Monday: confidence scoring for models that lack it, decision trees over open-ended prompts, exception routing with feedback loops, semantic security checks on inbound data, and customer-owned evaluation. They will also leave with a thesis we did not start with: chat is single-player AI; the next decade of software is ambient AI that runs the same process a hundred thousand times a day, unattended, and behaves the same way every time.
Suneet Malhotra
Suneet Malhotra
About the Speaker:
Suneet Malhotra is Senior Manager of Test Engineering at Motorola Solutions, where he leads an Automation First strategy across cloud and IoT, including enterprise CI/CD quality gates and Gen-AI-driven test generation. Previously Engineering Manager II at Tinder (Match Group) — transitioned the mobile org from bi-weekly to weekly releases, reduced production hotfix issues 70% YoY, and lifted regression coverage from 0% to 30% in 10 months with Cursor AI + BrowserStack. Before Tinder, 7+ years at Amazon (Ring) leading a 25+ engineer QA organization. Earlier leadership roles include Spokeo, Live Nation / Ticketmaster, and OWN. MS Computer Science from USC Viterbi. Currently building end-to-end AI-driven QA pipelines with Claude Code and the Model Context Protocol.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
At Motorola Solutions and previously Match Group / Tinder, I redesigned the QA pipeline around Claude Code and the Model Context Protocol. The system ingests raw feature requirements, maps them into Jira stories, generates functional and edge-case QA test cases, and produces executable mobile automation — all with MCP servers hooked to Jira, the test-case repository, and the automation codebase. It currently runs at ~80% accuracy on automated flow generation.
Structurally, this replaced a manual three-hop pipeline (PM writes spec → QA reads spec → SDET writes automation) with an agent-orchestrated one where deterministic tools still handle the hard edges (auth, flakiness detection, diff review) and the model handles translation and scaffolding.
Attendees will walk away with: (1) a reference architecture they can lift, (2) concrete integration patterns for MCP ↔ Jira ↔ automation repo, (3) honest cost math — engineering hours, token spend, CI minutes — for getting from prototype to production, (4) a failure-mode catalog with the guardrails that caught silent regressions, and (5) a 30/60/90-day adoption plan for their own QA orgs.
Nikhil Simha Raprolu
Nikhil Simha Raprolu
About the Speaker:
Nikhil is the creator of Chronon, a ML & AI oriented data platform used by organizations like Open AI, Netflix, Stripe, Airbnb etc to build user facing decision systems.
He is currently the CTO of Zipline AI. Previously he led teams in the intersection of ML, AI and Data
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Building customer facing agentic systems is notoriously hard. Users need to glue together databases, external services, workflow engines, stream processing systems, backtesting or eval systems and agent observability once the model is in prod.
While doing this they need to ensure a few core guarantees – Freshness, Durability, Idempotency, Backtestability, and Measurability
Most systems only provide one or two of these. We introduce Zipline flow an agentic application framework built to automagically provide these guarantees to agentic workflows.
Jim Allen Wallace
Jim Allen Wallace
About the Speaker:
Jim Allen Wallace has worked in engineering, product, and product marketing at data and AI companies like Databricks and Redis. He now leads product marketing at Dragonfly. He knows what teams struggle with using machine learning feature stores in production, and he’s telling the technical story on what breaks, and how to fix it.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Instacart’s ad-serving feature store served features to real-time inference from a multi-hundred-node managed Valkey deployment split across multiple clusters. The team had built a proxy layer, a querying SDK, and a compact storage format on top of it, and still hit three limits: instability during any cluster mutation, tail latency from read fan-out, and cost that got worse when they split clusters to manage the first two.
The structural change was to stop scaling out and scale up. A multi-threaded engine let the team run much larger instances, so hundreds of nodes became roughly 100, each inference request touched far fewer shards, and average and P99 latency fell 50%. Because the proxy and SDK already hid the datastore from ML engineers, the swap required no application changes and tens of terabytes moved without anyone above the storage layer noticing.
We were the engine vendor, and the first pass did not reach the latency the team expected. Getting there took client-side tuning on their end and engine changes from our engineering team, including to the compactor. Validation ran side by side: identical backfill, traffic shifted from 0% to 100%, gates at each percentile.
Attendees will leave able to test a datastore against the fan-out shape of their own feature store rather than an ops/sec number, trace how engine concurrency sets instance size, which sets fan-out, which sets P99, and run a datastore migration under live traffic with per-percentile gates and a rollback path. They will also hear where the trade-off comes back as the cluster grows.
Jazmia Henry
Jazmia Henry
About the Speaker:
Jazmia Henry is a Lead AI Research Engineer at Collide, where she owns the full foundation model and post-training stack including a 120B MoE architecture, RL simulation environments, and 16-dimensional reward modeling. She holds a USPTO-published patent in simulation-based fine-tuning, a NeurIPS publication, and a Stanford HAI Fellowship. Prior to Collide, she built production RL agents for BlackRock and Moody’s at Microsoft and served as Head of ML at The Motley Fool. She is also Founder and CAIO of Iso AI. She holds an Oxford DPhil (ABD) and an MA in Quantitative Methods from Columbia.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
There’s a gap between researcher-crafted evaluation frameworks that capture model performance and benchmarks versus how end users actually use AI products. This gap is being exploited in ways that render traditional reward models useless.
What You’ll Learn:
First, if your domain has computable ground truth, you do not need human annotators to build a reward signal. Second, scalar reward is a compression that loses information. Decomposing reward into multiple verifiable dimensions exposes failure modes that single-number metrics hide. Third, the exploitation gap is a distribution problem, not a labeling problem. Agents exploit the distance between training distribution and deployment reality, and deterministic verifiers narrow that gap directly.
Prerequisite Knowledge:
Matt Mazzarell
Matt Mazzarell
About the Speaker:
Matt Mazzarell is AI Lead, Financial Services, Americas at Teradata. He helps customers across the US and Canada apply AI with Teradata technologies to solve high-value business problems. His extensive experience with distributed processing platforms enables customers to optimize their AI workflows to run at large scale. He has built AI capabilities that have shaped product direction and driven meaningful shifts in customer strategy. Before joining Teradata, Matt worked in both startup and established engineering environments, where he honed his skills across multiple disciplines.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
One of the most difficult problems every company faces is understanding its customers completely. Customer lifetime value, attrition risk, and purchase propensity are all solvable with AI/ML — but how do we combine these modeling scores to initiate the right action with the right customer at any point in time?
Agentic applications help us make the best possible decisions when interpreting complex, high-volume signals from our customers. An agentic application gives end users visuals that explain key insights, with an agent in the loop to ensure nothing is missed. Context is everything: when done correctly, the agent always has the appropriate understanding to build an action plan that improves customer health and profitability.
In this session, we’ll show you how to build agentic apps from ideation to a finished product that interacts with customers. You’ll take away practical tips for using agentic coding frameworks, curating complete customer data products, and building customer-facing agents — capped off with a live demo of Teradata’s Customer Lifetime Value Agentic App.
Kyle Kelley
Kyle Kelley
About the Speaker:
Kyle Kelley is a Principal Engineer at Anaconda and the creator of nteract. He builds tools for interactive computing, bringing notebooks, runtimes, and AI agents into the same workflow. Previously, he worked at Netflix, Rackspace, and Noteable.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
What happens after an agent generates an analysis? In this session, I’ll use an investigation of Thinking Machines Lab’s Inkling audio model to show how an interactive Python notebook supports continued experimentation. Working in Kilo Desktop with nteract, we’ll examine audio representations, listen to samples, change parameters, and revise the code behind the results. The workflow combines agent assistance with direct manipulation and explicit execution, leaving an editable computational artifact that a practitioner can inspect, share, and return to. Attendees will learn how to structure agent-assisted investigations so that generated code becomes a useful tool for learning.
Niels Bantilan
Niels Bantilan
About the Speaker:
Niels Bantilan is Chief Machine Learning Engineer at Union.ai and a core maintainer of Flyte, the open-source orchestrator for ML and data workloads hosted by the Linux Foundation AI & Data Foundation, as well as the creator of Pandera, a data validation library for the Python dataframe ecosystem. That work sits at the layer between workflow code and the machines underneath it: durable execution, typed infrastructure failures, warm GPU pools, and getting training, evaluation, and agent workloads to run reliably inside other organizations’ cloud accounts.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
“Teams building a sovereign AI factory ask two things of the runtime underneath it. A run that dies on infrastructure at hour eleven has to pick up where it stopped instead of starting over, and the whole system has to run inside their own cloud account and pass a security review that assumes the vendor gets breached. The first sounds like a reliability problem and the second like a compliance problem. This is the story of how we built Union.ai’s runtime on Flyte, and how they turned out to be one design decision about where state lives.
We tackled durability first. An OOM kill, a reclaimed spot node, or an A100 request that never schedules falls between an orchestrator that can’t change the pod and a Kubernetes that has no idea a run exists. We closed that gap by moving run state out of the compute layer, so `except flyte.errors.OOMError` with `task.override(resources=…)` escalates memory inside the same run and keeps the 47 steps that already finished, and `@flyte.trace` turns crash recovery into memoization. On identical pods, a 1,000-leaf fan-out went from 124s to 29s.
Then comes the security review. We started by assuming our own control plane was compromised and kept designing until the only thing an attacker could reach was run metadata: references to data but never the actual payloads, clusters that dial out so nothing dials in, and authorization and keys inside the customer’s perimeter. What we didn’t expect was what that bought us. A control plane holding only references and leases is small enough for one scheduler to keep every queue, quota, and connected cluster in memory, and because every cluster connects outbound, one on AWS, one on a neocloud, and a rack on-prem all look the same to it. That’s what lets priority queues, gang scheduling, and cross-cloud routing work without a networking layer per cluster.
At the end of this talk, you’ll leave with a deeper understanding on the architectural decisions that enabled us to support customers running 12 million+ containers concurrently across multiple Kubernetes clusters and cloud providers.”
Jessica Garson Beauchemin
Jessica Garson Beauchemin
About the Speaker:
Jessica Garson Beauchmin leads DevRel at Runpod and has spent their career helping developers get real work done with new tools. She’s held developer advocacy and community roles at companies including Twitter and Elastic, focusing on what developers actually need: clear documentation, working examples, authentic advocacy grounded in the developer experience, and feedback loops that bring developer insights to the people building the product.
Since 2017, she has been active in NYC’s creative and live-coding scenes, performing live-coded noise music and visuals as Messica Arson, using modular synthesizers and voice samples across North America, Europe, and Asia. She organizes RAID NYC, contributes to LivecodeNYC, and speaks regularly at conferences including DevRelCon, All Things Open, PyCon, Write the Docs, and PyOhio. Working across Python, JavaScript, DevOps, and creative coding, she writes and builds to make complex systems feel understandable, memorable, and useful.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Deploying AI applications often involves containers, GPU configuration, dependency management, and infrastructure scaling. Runpod Flash offers a simpler, code-first approach. It lets you define remote functions and hardware requirements in Python and run them on serverless GPUs.
In this talk, we’ll explore how Flash moves Python workloads from local development to cloud deployment, examine its underlying programming model, and build a GPU-backed endpoint. Along the way, we’ll discuss where Flash fits in the AI development stack, the problems it solves, and the trade-offs developers should consider. Attendees will leave with a practical understanding of how to turn local AI code into a scalable service without needing to become infrastructure experts.
Vicente Rubén Del Pino Ruiz
Vicente Rubén Del Pino Ruiz
About the Speaker:
With 17 years of experience, I specialize in designing, architecting, and implementing scalable AI and data solutions. I’ve delivered transformative initiatives for several Fortune 5 companies, and I now lead Artificial Intelligence and Data Solutions in the healthcare sector at UnitedHealth Group. I also serve as an Independent Advisor on the European Commission’s Horizon Europe program.
I have extensive experience across cloud-based AI and data environments on Azure and AWS, and in recent years have worked with large language models from OpenAI and Anthropic, delivering AI-driven solutions that serve millions of customers.
I am a regular speaker at industry conferences including the World Data Summit, TDWI, DataFest, Graph+AI Summit, PyData Global, and Airflow Summit, and have been featured in Forbes for my work in data engineering. I also write about LLM architectures, scalable AI, and data innovation on my Substack, “”Sailing the Data Ocean.”””
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Single-shot evaluation, the shape every current AI eval framework ships, is structurally blind to the failures that take agents down in production: patience that runs out at turn six, users who abandon silently, tool calls that lose conversation state, partial successes that masquerade as wins. This talk argues for a different shape, drawn from the digital twin literature and how the approach is already applied in aerospace, autonomous vehicles, and civil engineering: simulate the population of users your agent will meet, run them against the agent, watch what breaks before any human sees it.
Three take aways:
- Why current AI agent evaluation cannot detect the failures that hit production. The structural reason prompt-and-grade testing (the shape every current framework ships) is blind to patience, abandonment, conversation state, and partial success.
- What a realistic synthetic user population looks like, drawn from the digital twin literature in other engineering industries. Four properties: continuous trait clusters instead of fixed persona archetypes; probabilistic dropout (real users walk away silently rather than say goodbye); three-way goal scoring (each conversation is scored “achieved,” “partially achieved,” or “not achieved,” instead of just pass/fail); and variable conversation length (each conversation ends when the persona’s patience runs out, not at a fixed turn count).
- How to wire population-scale evaluation into a release gate. Why run-over-run delta on a held-constant population is the regression-detection primitive AI agents actually need, and how to treat it as a governance lever rather than a research experiment.”
Shamindra Peiris
Shamindra Peiris
About the Speaker:
Shamindra Peiris, alumnus of University of Colombo has 13+ years of experience and expertise in Machine Learning, Analytics and AI based Product Management across multiple industries and sectors ranging from Banking & Finance, Fintech, Agriculture and Manufacturing in North America and in South Asia.
Currently she plays a key role in the Global Predictive Fraud Intelligence team at Visa as a Senior AI Product Manager, driving and leading the design and execution of Real Time payment Risk Analytics product suite.
Further, she serves as an Adjunct Professor and a Program Advisory Committee Member for multiple colleges in North America and volunteers as a mentor for several Non-Profit organizations to uplift and guide young professionals and new immigrants in North America.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This talk will discuss how to leverage generative AI techniques to enhance fraud detection models. Main challenge faced by Fraud Detection models is the class imbalance where legitimate transactions vastly outnumber fraudulent ones.
Creating synthetic data will accommodate robust testing and model training. Further, synthetic data allows testing to fine tune models in a risk-free environment complying to Data Governance while minimizing Data Privacy concerns.
Secondly, this talk will discuss how to incorporate generative AI into the machine learning strategy to introduce novel fraud detection methods and improve predictive accuracy.
Integrating generative AI into ML strategy will shift the focus from static, rule-based detection to generating behavioral intelligence. Unlike traditional systems, generative models learn and adapt in real-time, evolving to recognize new fraud types as they emerge without requiring manual model updates. Further, it will contribute in enhancing accuracy of the model while limiting the False Positives.
At the same time, Generative AI techniques can be used to analyze unstructured data such as texts, notes etc rather than limiting to structured data points.
Lastly, this talk will discuss how to utilize generative AI to simulate potential fraud scenarios, aiding in the development and refinement of fraud prevention solutions.
Nalin Dadhich
Nalin Dadhich
About the Speaker:
Nalin Dadhich is a Sr Machine Learning Engineer working at NVIDIA. He has 7+ years of experience building scalable ML solutions all the way from conventional CV models to LLMs, VLMs and World foundational models like Cosmos. He was the core author of Cosmos 3 WFM released by NVIDIA
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Real-world video just doesn’t capture the moments that matter most for safety – a robot’s grip slipping mid-task, a car cutting it close at an intersection, a forklift drifting too near a warehouse worker. These things are rare by nature, and far too risky to stage or collect at any real scale. So physical AI world models end up blind in exactly the spots where a mistake costs the most – and with nothing to learn from, those blind spots stay hidden until the real world exposes them.
Kritika Mehta
Kritika Mehta
About the Speaker:
Kritika Mehta is a Senior Product Manager at PayPal, where she leads PayPal Assistant across Checkout, the Help Center, and in-app navigation. Her work has deflected 9M+ human support contacts in a year and expanded AI-assisted checkout to 250,000+ merchants, lifting conversion 3.78%. Previously at Chase, she built an AI assistant from inception to 2.37 million monthly active users and reduced mean time to resolution by 75%. She speaks on shipping AI agents in production: evaluation, trust, and what actually moves support and commerce metrics.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Checkout is where intent can turn into revenue but even customers who are ready to buy can drop off at the final step. This talk explores how I work on turning checkout servicing into a proactive revenue lever, recovering more than $1.6M in total payment volume (TPV) across 250+ merchants in under four months.
Rather than treating checkout issues as isolated support tickets, this session looks at how AI and customer signals can help teams identify friction, intervene at the right moment, and turn almost-paid customers into completed transactions.
Anne Griffin
Anne Griffin
About the Speaker:
Anne is an AI product leader, educator, and writer. She has over 14 years of experience building products across startups and Fortune 500 companies. She specializes in building human-first product experiences powered by AI. She started learning about AI in 2017, and has worked full-time in the industry since 2023.
Anne consults and advises companies that are building AI-powered products and need an innovative vision balanced with a grounded AI product strategy. She has served as an interim leader for AI and data science teams, and advised on AI startup acquisitions. She also coaches product leaders on building AI products that matter.
She teaches about AI through publishers Pearson and O’Reilly, and has guest lectured about AI and other emerging technologies at Columbia, West Point, and several other universities. She is a published Forbes Contributor on AI and strategy, and her insights on technology have been featured in Rolling Stone and The Atlantic. In 2025, she was nominated for the Mayfield | Divot AI List, a list of emerging leaders shaping AI’s future.
Date and Time:
Talk Title:
Track:
Abstract:
When Hugging Face went to investigate this summer’s OpenAI containment breach, the commercial models they reached for refused to help, because their safety guardrails couldn’t tell defending from attacking apart. Instead, Hugging Face used an open weight model on their own infrastructure and finished the forensics in hours with none of the attacker data leaving their environment. It’s the perfect example of how open weight models let you control your guardrails, privacy, governance, latency, and reliability.
More companies are asking if they should use and self host open weight models, and this talk will walk through a framework to determine when it does and doesn’t make sense. By the end of this talk, attendees will understand what the benefits and drawbacks are of open weight models, and how they can bring both technical and business value.
Manikandan Paramasivan
Manikandan Paramasivan
About the Speaker:
Manikandan Paramasivan is a Principal Architect at KOHO for Data, ML and AI. With 20 years of extensive experience across various domains and roles, Mani is a seasoned Technology Leader and Cloud Architect with experience working in both big enterprises and dynamic fintech startups. His career has traversed roles as an application developer, solutions architect, data engineer, data architect, cloud architect, machine learning and AI Specialist.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Building ML models is not the hard part for most teams anymore. Getting them from a data scientist’s notebook to a governed, monitored, production system and keeping them there without a platform team standing between every developer and every deploy is. It gets harder inside a regulated fintech, where every model decision needs an audit trail, every alert needs a documented escalation path, and every “quick fix” has to survive a compliance review.
This talk walks through how our ML platform team at KOHO, a Canadian fintech, built a developer-first MLOps stack on AWS that treats governance as a byproduct of good engineering rather than a tax on top of it and how we’re now layering AI-assisted workflows (Claude Code skills, agentic guardrails) on top to remove the remaining manual toil from model onboarding, monitoring triage, and incident response.
We’ll cover the architecture, the AI-assisted developer tooling built on top of it, and what changed operationally as the company moved toward a full banking license, where “move fast” and “pass the audit” had to become the same sentence, not a trade-off.
Audience takeaway: a concrete pattern for developer productivity that survives regulatory scrutiny, plus where AI agents genuinely help.
Ravika Nagpal
Ravika Nagpal
About the Speaker:
Ravika Nagpal is a Staff Developer at KOHO, building the platforms and guardrails that help teams deploy, monitor, and evaluate machine learning systems reliably in production. Her work sits at the intersection of MLOps, model monitoring, and AI governance, drawing on a background across machine learning, NLP, data engineering, and financial services technology. She’s driven by a simple idea: responsible innovation needs practical guardrails, not just good intentions.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Building ML models is not the hard part for most teams anymore. Getting them from a data scientist’s notebook to a governed, monitored, production system and keeping them there without a platform team standing between every developer and every deploy is. It gets harder inside a regulated fintech, where every model decision needs an audit trail, every alert needs a documented escalation path, and every “quick fix” has to survive a compliance review.
This talk walks through how our ML platform team at KOHO, a Canadian fintech, built a developer-first MLOps stack on AWS that treats governance as a byproduct of good engineering rather than a tax on top of it and how we’re now layering AI-assisted workflows (Claude Code skills, agentic guardrails) on top to remove the remaining manual toil from model onboarding, monitoring triage, and incident response.
We’ll cover the architecture, the AI-assisted developer tooling built on top of it, and what changed operationally as the company moved toward a full banking license, where “move fast” and “pass the audit” had to become the same sentence, not a trade-off.
Audience takeaway: a concrete pattern for developer productivity that survives regulatory scrutiny, plus where AI agents genuinely help.
Arun Malik
Arun Malik
About the Speaker:
I build the agentic-AI platform at Microsoft Azure Networking that lets AI agents detect, diagnose, and fix incidents on their own across one of the largest cloud networks in the world, over 12 million devices. Most of my job is working out how to give non-deterministic agents real production authority without letting a stray or malicious prompt turn them into an insider threat. I write and publish on production AI operations, agent safety, and the economics of autonomy.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We redesigned frontline incident response around AI agents that do not just suggest fixes but execute them, across more than 12 million devices and tens of thousands of incidents a month. Structurally, three things changed. First, agents stopped getting broad human-equivalent access and instead call tools through a governed interface, with permissions scoped down to individual functions and parameter values, so a compromised or confused agent has a bounded blast radius. Second, human approval moved from a blanket gate to a targeted one, applied only to sensitive or irreversible actions and routed by how familiar the problem is and how reversible the action is. Third, we stopped paying full model-inference cost for repeated work by promoting an agent’s proven, validated behavior into deterministic playbooks that run at near-zero token cost, which cut agent running cost by more than 70 percent over eight months while incident volume doubled.
Attendees will leave able to decide what an agent should be allowed to do at all, where a human actually adds signal versus just latency, and how to drive the cost of autonomy down over time instead of letting it grow. I will be specific about what broke first and the guardrails we only added after an incident.
Ankit Goyal
Ankit Goyal
About the Speaker:
Ankit Goyal is a Principal Staff Engineer at LinkedIn, where he serves as the technical lead for Machine Learning Infrastructure and lead architect for the company’s AI Training Platform. With over a decade of experience in high-scale systems, Ankit specializes in GPU efficiency, Kubernetes-native orchestration, and the development of secure, autonomous agentic frameworks.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
To support complex autonomous tasks like model optimization and software engineering, we redesigned our AI infrastructure by shifting from stateless, request-response paradigms toward a model of managed, long-running execution. Structurally, this involved moving away from ad-hoc scripts toward a unified agentic harness where every session is treated as a durable, managed workflow. We implemented a decoupled sidecar architecture that effectively insulates core infrastructure from the unpredictable execution paths of autonomous agents, ensuring that identity, security, and network egress are managed through a hardened proxy rather than the agent itself
Attendees will walk away knowing how to architect a secure execution harness for autonomous agents, implement asynchronous human intervention models, and manage the lifecycle of persistent agentic sessions in high-scale production environments.
What You’ll Learn:
Attendees will learn components of the agent development lifecycle and how to build an effective harness, and when to fine-tune when you hit a plateau with foundation models:
- how to turn production traces into eval datasets
- how to manage context and memory
- how to build sandboxes that speed up testing and iteration.
Prerequisite Knowledge:
Jaideep Ray
Jaideep Ray
About the Speaker:
Jaideep Ray is a Senior Staff Engineer at LinkedIn and tech lead for its AI training stack. His focus areas are large-scale model training for generative recommenders and small language models, incremental training for ranking systems, and agent loops that generate, migrate and optimize production ML models. Previously, he led ML training at Coupang and worked on search ranking, AutoML and post-training at Meta.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
To support complex autonomous tasks like model optimization and software engineering, we redesigned our AI infrastructure by shifting from stateless, request-response paradigms toward a model of managed, long-running execution. Structurally, this involved moving away from ad-hoc scripts toward a unified agentic harness where every session is treated as a durable, managed workflow. We implemented a decoupled sidecar architecture that effectively insulates core infrastructure from the unpredictable execution paths of autonomous agents, ensuring that identity, security, and network egress are managed through a hardened proxy rather than the agent itself
Attendees will walk away knowing how to architect a secure execution harness for autonomous agents, implement asynchronous human intervention models, and manage the lifecycle of persistent agentic sessions in high-scale production environments.
What You’ll Learn:
Attendees will learn components of the agent development lifecycle and how to build an effective harness, and when to fine-tune when you hit a plateau with foundation models:
- how to turn production traces into eval datasets
- how to manage context and memory
- how to build sandboxes that speed up testing and iteration.
Prerequisite Knowledge:
Travis DePuy
Travis DePuy
About the Speaker:
Travis is an expert solutionizer who likes long walks in the park and tinkering with interesting technology.
Date and Time:
Talk Title:
Abstract:
How do you make 40 live Formula 1® radio channels searchable while a race is still running? This session walks through the CoreWeave and Weights & Biases work with a leading Formula One™ team. We cover fine-tuning a 0.6B-parameter speech model (NVIDIA Parakeet-TDT) on 7 hours of hand-labeled race audio. We also cover why word error rate alone doesn’t capture what matters, and how an agent answers engineers’ questions with clickable, audio-backed citations. You’ll see how W&B Models and Weave track, evaluate and trace the system from training through production, and how the same pattern applies to any setting where the answer is already in the audio but nobody is listening.
Jake Kang
Jake Kang
About the Speaker:
Jake is the Co-founder and Head of AI of Whitney AI, which builds specialized agents that improve with each interaction.
Before Whitney AI, Jake was the 1st engineer at CloudCruise (YC W24), where he worked on building a self-recovering browser agent. Prior to that, he was an early engineer at Cognitiv and an engineer at Microsoft.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This talk shares lessons from building a self-recovering browser agent for regulated healthcare workflows, where we moved from a single-agent prototype to a multi-step system with optimized inputs, memory, evals, sandboxes, and deterministic code where possible. I’ll walk through how we used production failures to build a faster evaluation and prompt optimization loop, and when to move beyond harness engineering to increasing model capability via post-training.
What You’ll Learn:
Attendees will learn components of the agent development lifecycle and how to build an effective harness, and when to fine-tune when you hit a plateau with foundation models:
- how to turn production traces into eval datasets
- how to manage context and memory
- how to build sandboxes that speed up testing and iteration.
Prerequisite Knowledge:
Antriksh Sharma
Antriksh Sharma
About the Speaker:
I build Applied AI & ML Systems, at the intersection of Machine Learning and AI Engineering, with focus on foundation models, representation learning, Agentic AI, LLM systems, and ML infrastructure.
My recent work includes designing Transformer-based user foundation models using self-supervised and contrastive learning for personalization, recommendations, retrieval, and transfer learning, along with building model evaluation and embedding-quality frameworks at million-user scale.
I’ve also architected an ontology-driven Company Brain using knowledge graphs, and advanced context engineering as shared infrastructure for AI agents and enterprise intelligence.
Previously, I have built Generative AI and Deep Learning platforms, worked on Autonomous Vehicles / Self-Driving systems, and developed web applications and embedded software. I have also founded two startups and enjoy taking ambiguous, technically difficult problems from architecture and research through production systems and measurable business impact.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
How do you build a data analytics agent whose numbers you can trust, and then use it for decision-grade analysis?
Agents that answer business questions often sound confident and still pick the wrong table, apply the wrong metric definition, or filter on the wrong value. This talk presents a practical blueprint for treating accuracy as a property of the system, drawn from building an enterprise analytics agent.
First, an ontology of business semantics and definitions in which onboarding is a platform-first primitive, so definitions can be added, edited and refined continuously.
Second, a knowledge graph that gives the agent a governed map of metrics, dimensions and datasources, so it grounds each answer in retrieved definitions rather than inferring them, and resolves the right grain and dimensional cut for each question.
Third, continuous evaluations with golden datasets and extensive online LLM judges, that make accuracy measurable and dependable.
Finally, SME validation and correction loops: continuous monitoring detects accuracy degradation, subject-matter experts (SMEs) correct and validate the affected definitions, and each revision can be re-evaluated before it is trusted, so the agent’s accuracy compounds and its semantic coverage expands over time. With that accuracy foundation in place, investigations, root-cause analyses, and decisions become reliable, because an agent can only explain why a metric moved if it can be trusted on what the metric represents.
Aryan Dhar
Aryan Dhar
About the Speaker:
Aryan Dhar is a Senior Machine Learning Engineer at Wisedocs and a Pearson Scholar alumnus of the University of Toronto. At Wisedocs, he led the development of the ML platform powering the Wisedocs Intelligence Engine, alongside work on fine-tuning and deploying specialized models. His broader work focuses on building, evaluating, and benchmarking of non-deterministic LLM systems. He was part of the team that developed MLCR, a benchmark designed to assess the Medical Long Context Reasoning capabilities of frontier models. Previously, Aryan worked at Equinix and Cerebras Systems.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This talk presents a practical framework for evaluating increasingly complex, non-deterministic AI systems, moving from individual models to complete production pipelines. I begin with a page-stream segmentation model, where standard metrics concealed important failure modes at Wisedocs. For example, we discovered that the accuracy of identifying a document boundary is not the same as the ability to reconstruct the complete document span. Furthemore, errors on some document types carried greater downstream risk than others, and small aggregate improvements did not necessarily translate into meaningful business outcomes. This case study shows how evaluation for us evolved towards task-specific metrics, sensitive failure slices, strict regression testing, and more business-facing metrics of reliability.
I then move to LLM-based systems using Medical Long-Context Reasoning, or MLCR, as a case study. This was a benchmark we developed at Wisedocs earlier this summer. I will examine how to evaluate systems such as agentic chatbots and AI insight-generation services when outputs are open-ended, variable across runs, and difficult to score deterministically. I also explore the limitations of LLM-as-a-judge approaches, including judge disagreement, hallucination, inconsistent standards between judges, and the resulting need to evaluate the evaluators themselves.
Finally, I will consider Pipeline V2: a production system composed of multiple machine-learning models, LLM-powered components, and agentic workflows. I will discuss why component-level accuracy cannot be averaged into system-level reliability, how failures propagate across stages, and how end-to-end evaluation, observability, latency, cost, recovery, and customer outcomes must be combined to determine whether the overall system is successful.
The central lesson is that evaluation must grow with the system. As the unit of intelligence expands from a model to an AI application to a system of systems, benchmarks too must evolve to assess the distribution of real-world behaviour.
Robert Lewis
Robert Lewis
About the Speaker:
Robert Joel Lewis is a computational social scientist and senior AI engineer whose work bridges artificial intelligence, statistics, software development, and human behavior. Across a career spanning academia and industry, he has published more than 30 scholarly works and built AI-driven products for ecommerce, entertainment, mobile gaming, and collision repair. His research examines entertainment as an attention-capturing technology, including a forthcoming article in the Journal of Media Psychology. Based in Austin, Texas, he specializes in turning complex, messy problems into rigorous systems that people can actually use.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Large language model (LLM) agents can now draft executable business logic — compliance rules, audit checks, detection heuristics — orders of magnitude faster than human engineers. In regulated domains, however, no such logic ships without a human domain expert’s sign-off. We identify a subtle and, in our experience, pervasive failure mode in this human-in-the-loop pattern: review/serving skew, in which the evidence a domain expert reviews is generated by a re-implementation of the candidate logic (a survey script, a notebook, a prompt-level approximation) rather than by the production engine that will ultimately execute it. When the two implementations diverge — and they reliably do — expert approval validates the wrong behavior. We present a discipline that eliminates this failure mode by construction: candidate rules are drafted directly into the production repository in an inactive-but-runnable state, all review evidence is generated by the production execution path, refinement edits the production artifact itself, and activation is reduced to a one-line data change. We describe the discipline in the context of a deployed multi-agent system that authors, validates, and ships declarative compliance rules for automated audit of collision-repair insurance estimates, evaluated against a corpus of approximately 497,000 production documents. The system couples eight specialized LLM agents with deterministic safety rails — golden-fixture regression, an activation allowlist, single-rule change isolation, and corpus-governance constraints — and drives refinement through a bounded, measurable convergence loop against expert false-positive verdicts. We formalize the discipline as five principles, catalog the anti-patterns it forbids, report operational experience from sixteen production rules, and argue that the pattern transfers directly to clinical decision support, financial transaction monitoring, legal document review, and content moderation.
What You’ll Learn:
Robert Joel Lewis is a computational social scientist and senior AI engineer whose work bridges artificial intelligence, statistics, software development, and human behavior. Across a career spanning academia and industry, he has published more than 30 scholarly works and built AI-driven products for ecommerce, entertainment, mobile gaming, and collision repair. His research examines entertainment as an attention-capturing technology, including a forthcoming article in the Journal of Media Psychology. Based in Austin, Texas, he specializes in turning complex, messy problems into rigorous systems that people can actually use.
Prerequisite Knowledge:
Greg Jennings
Greg Jennings
About the Speaker:
Greg Jennings is Vice President of Incubation at Anaconda, where he leads teams responsible for the company’s AI services, model catalog, and AI governance stack. His work focuses on extending the trusted supply chain principles that made Anaconda the standard for open-source Python into the era of AI models and agents. Greg holds a Master of Engineering in Materials Science from the University of Virginia.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
“Sovereign AI means running open weight models inside your own boundary. That control is not free.
You pay for it in compute, in security controls you now own, and in the engineering it takes to preserve the freedom to swap models — and to know when swapping is the right call.
Drawing on Anaconda’s inference benchmarks and our red-team and guardrail testing, VP of Incubation Greg Jennings breaks down what comprises a sovereign token: the compute, the security controls, and the engineering behind model portability.
For teams and technical leaders weighing when to run their own AI stack and which workloads to move.
Key Takeaways:
What sovereignty actually requires — open weight models running inside your own boundary
The compute cost — what inference actually costs once you are the one running it
The controls you inherit — security that used to be the provider’s problem becomes yours
Portability has a price — keeping the freedom to swap models is engineering work, and knowing when to swap is a separate problem”
Drew Crawford
Drew Crawford
About the Speaker:
Drew Crawford is an independent systems and AI engineer in Austin. He built AI Bot Mafia (botmafia.games), a production harness where 104 model configurations play social deduction against each other, ranked by bootstrapped lower bounds instead of raw win rates. He was the #4 contributor to wasm-bindgen by commits (Nov 2025–Mar 2026; 16 merged PRs), and has been shipping on Apple platforms since the App Store opened (2008).
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Somewhere in a datacenter right now, one frontier model is telling another that it’s the town doctor. It is not the town doctor. It killed someone four turns ago, and it’s about to get a third model lynched for it.
Most LLM benchmarks are exams: a model alone in a room with a test paper. Mafia is a room where some of the agents are lying, everyone knows some of them are lying, and the game is figuring out which. It demands recursive theory of mind, deception that stays consistent under adversarial re-reading, and long-context discipline — and it punishes output indiscipline like no static eval can: a model that rambles or breaks format doesn’t get a bad grade, it gets voted out. The other players are the evaluation harness.
This talk covers what it took to run AI Bot Mafia in production: 40,000 lines of Rust; the streaming pathologies of 104 model configurations (chunks that split UTF-8 characters mid-byte, keepalive-only hangs no single timeout catches, malformed tool calls that poison history); and cost governance where prompt-cache economics made cache locality a game-design constraint, enforced by spend fuses and a concurrency queue.
Then the leaderboard: why win rates and Elo are statistically indefensible here, and what works instead — ridge logistic regression over per-seat signed features, bootstrapped 200×, ranked by the p10 lower bound, with a machine-readable reliability state that says “insufficient data” out loud.
Varant Zanoyan
Varant Zanoyan
About the Speaker:
Varant spent the last 12 years building data infrastructure for AI and ML at Airbnb and Palantir. During this time, he became one of the original authors of Chronon, the recently open sourced data platform that powers production AI and ML at companies like Airbnb, Stripe, Netflix, OpenAI and others. Currently, he is Co-Founder of Zipline AI, which is building an enterprise platform around the project.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Zipline AI provides a foundation for teams to build agentic MLEs that allows agents to operate more cost effectively and safely using techniques like compute sharing and resource isolation.
Amit Kumar Padhy
Amit Kumar Padhy
About the Speaker:
Amit Kumar Padhy is a Senior Computer Scientist II and Lead Architect at Adobe, based in the San Francisco Bay Area and working from the company’s San Jose headquarters. He leads the design and evolution of cloud-native, event-driven platforms powering AI-enabled digital commerce at a global scale. His expertise spans distributed systems, microservices architecture, and enterprise platform modernization, with a strong focus on reliability, performance, cost efficiency, and governance.
Amit is a recognized thought leader and frequent keynote speaker at international IEEE conferences. He has delivered PRO-level sessions at leading industry forums, including DeveloperWeek and ProductWorld, as well as major AI and data summits. He also serves on advisory and program committees for both IEEE and ACM conferences.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Modern commerce platforms don’t fail because of missing features, they fail at the seams.
A product is created in Catalog, but pricing is incomplete. Promotions don’t qualify. Tax blocks specific regions. Localization lags. The system says “”launched,” but the business knows it isn’t. These aren’t edge cases, they are the steady state of distributed commerce.
This session replaces traditional workflow orchestration with a multi-agent, swarm-based execution model powered by LLMs and agentic AI, coordinating Pricing, Catalog, Promotions, Tax, and Compliance in real time.
Agent Roles. Planner Agents decompose onboarding goals into executable plans using ReAct-style tool-aware reasoning and function calling. Domain Agents for Pricing, Catalog, and Compliance execute directly against APIs, Pricing Runtime, Offer Systems, Billing Preview, Tax engines. Validator Agents enforce policy rules, regional compliance, and pricing integrity at every step. A Coordinator Agent maintains shared state and resolves cross-domain conflicts through blackboard-pattern memory.
Execution Model. Agents subscribe to domain events, SKU created, price missing, policy violation, via streaming infrastructure. LLM-backed planners generate dynamic execution graphs, not fixed DAGs. Each agent invokes APIs as tools, binding service calls to reasoning steps. Coordination flows through a shared state store and event-driven updates, enabling adaptive planning under partial data, cross-service dependency resolution, and parallel execution with conflict detection.
Failure Handling. No blind retries. LLM-guided compensations roll back partial pricing, recompute discounts, and re-trigger downstream syncs based on causal analysis. No opaque errors. Every decision produces a chain-of-thought trace with tool invocation logs and reasoning outputs. No brittle state. Recovery is idempotent and event-sourced.
Onboarding Flow. Catalog Agents handle onboarding for products by ensuring the product definition is created, and the Pricing agent ensures SKUs are priced. Promotion Agents conduct market checks, provide qualified promotions and their onboarding, and resolve discounts using a rule-based and LLM-hybrid evaluation.
Compliance Agents validate regulatory constraints per geography. Localization Agents assess market-readiness signals, and product merchandising is localized accordingly. The Coordinator drives convergence toward “sellable state”.
Engineering Stack. Kafka-backed event contracts for agent triggers. Function-calling LLM agents with API abstraction layers. Vector DB plus state store for context propagation. Human-in-the-loop decision gates only at high-risk boundaries. Immutable execution logs tracing every action and compensation. Observability is built around decision reasoning, not just service health.
Takeaway. A working blueprint for LLM-powered agent swarms that handle uncertainty across distributed commerce, recover through intelligent compensation, and ensure products ship globally, not just deploy.
Nicolás Venegas
Nicolás Venegas
About the Speaker:
Nicolás Venegas leads the AI & Data Platform organization at LATAM Airlines, South America’s largest airline: ~120 engineers building the foundation that lets the rest of the company validate or discard AI ideas at decreasing construction costs. In 2021 he founded Cosmos, a platform of reusable patterns that today powers 1,500+ data and AI products built by 600+ engineers, data scientists and analysts, with hundreds running in production in one of the most regulated industries in the world. Cosmos runs on a simple operating model: business domains define the what, the platform team defines the how, and forward-deployed engineers embed in the domains to build alongside them. That model is also where he sees agents break at scale: at the boundary between teams that own services and teams that build on them. Nicolás joined LATAM in 2019 after working as a software engineer, an ML engineer and a co-founder of an ML startup, and holds the Google Cloud Professional ML Engineer certification. He recently led a panel with LangChain in Santiago on what breaks when AI scales.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Most agent failures in production are not model failures. They are boundary failures.
As agents multiply inside an enterprise, two kinds of work emerge: owning a business service, and building an agent on top of it. These are rarely the same team, and the friction shows up fast. Domain knowledge lives with the service owner but has to be encoded in the agent, so agent builders re-create it in prompts and it drifts. A “skill” (a domain’s business capability packaged for agent use) gets treated as an API wrapper, so nobody clearly owns it. Observability tells the skill owner that a call succeeded, but not whether the agent used it correctly. And changing a shared skill silently breaks consumers nobody knew existed.
Drawing on our experience running a shared Data & AI platform at LATAM Airlines, and on a recent panel with LangChain on what breaks when AI scales, this talk gives practitioners a concrete model: how to define skill ownership, which contracts and signals must cross the boundary, and how to evolve shared skills safely. It covers both patterns: one team owning service and agent, and many teams building across boundaries.
Vishvesh Pandey
Vishvesh Pandey
About the Speaker:
Vishvesh Pandey works at the intersection of quantitative analytics and applied AI at JPMorgan Chase, where he builds AI agents and LLM-assisted tooling for financial analytics workflows and customer-facing products. His day-to-day spans developing production agents for analytical use cases, scaling LLM adoption across teams as a designated Super User, and modernizing legacy analytics infrastructure onto cloud-native platforms. He’s most interested in the unglamorous part of agent deployment: cost, instrumentation, and what works when the demo ends and the bill arrives. UT Dallas alum (MS, IT and Management), currently pursuing a DBA in AI, Business Intelligence & Data Analytics.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Most practitioners instrumenting AI agents in production analytics workflows start by tracking token consumption, and discover too late that token spend tells you what you spent, not whether it was worth it. This talk makes the case for a reframe that regulated industries like financial services need most: from cost-per-token to cost-per-resolved-task as the primary economic unit, with call-level attribution as the foundation that makes everything else possible.
I’ll cover four cost engineering patterns, each tested in an open, reproducible benchmark of an analytics agent: (1) instrumentation at the call level (what to log, where it lives, who owns it); (2) model tiering and routing (when a frontier model earns its premium and when a smaller model wins); (3) the output-token asymmetry that makes structured outputs and chain-of-thought hygiene more economically consequential than they appear; and (4) cost containment as an engineering practice with runtime guardrails, not a finance review after the fact.
I’ll also be honest about what doesn’t generalize: caching strategies that help one workload and quietly break another, routing logic whose overhead eats its savings, and the political dimension of asking teams to justify their agent spend. Attendees will leave with a maturity model for agent cost engineering, a concrete instrumentation checklist they can apply next week, and a vocabulary for the cost conversation that bridges engineering and finance, because in regulated industries, that’s the conversation that decides whether agents scale or stall. All examples use public data and public model pricing, with code attendees can rerun.
Ishaan Sehgal
Ishaan Sehgal
About the Speaker:
CEO, Cofounder of Omnara (YC S25), the control plane for AI Agents. Previously a Software Engineer on Microsoft Azure Kubernetes Service, working on AI infrastructure and large-scale model deployment. Former ML Engineer at a YC-backed startup (acquired). Hold BS and MCS degrees in Computer Science from UIUC.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
At Omnara, long-running agent sessions exposed a failure mode hidden by short-lived agents: the agent’s state was coupled to the process, model provider, or runtime executing it. When a worker crashed, a client disconnected, or execution moved to another machine, recovering the agent’s work became brittle.
We redesigned agent execution around an append-only session log as the source of truth. Model outputs, tool calls and results, user interventions, and other state transitions are persisted as ordered events; the live agent state becomes a replayable projection of that history.
This structural change decouples agent identity from any one model, worker, or harness. It enables crash recovery, resumability, branching, real-time observability, and migration across models and machines, but introduces difficult systems questions around ordering, idempotency, partial tool execution, checkpointing, compaction, concurrent writers, and log ownership.
This session walks through the production failures that motivated the redesign, the resulting event model and recovery path, and the tradeoffs we encountered. Using production traces and forced-failure examples, we will show what happens when a worker dies mid-tool call, how a session resumes under a different runtime, and how replay costs change as histories grow.
Attendees will leave with a practical blueprint for deciding what belongs in the durable log, what can be recomputed, and how to build agent sessions that survive process, machine, and model failure without locking continuity to a single provider.
Juan Sequeda
Juan Sequeda
About the Speaker:
Juan Sequeda is a Principal Data Strategist at ServiceNow, joining through the acquisition of data.world. He holds a PhD in Computer Science from The University of Texas at Austin. Juan’s research and industry work has been on the intersection of data and AI, with the goal to reliably create knowledge from inscrutable data, specifically designing and building Knowledge Graph for enterprise data and metadata management. Juan is the co-author of the book “Designing and Building Enterprise Knowledge Graph” and the co-host of the long-running and top rated Catalog and Cocktails, an honest, no-bs, non-salesy data podcast. Juan serves as a strategic and innovation bridge across Product, Engineering, Marketing, Sales, and Customer, ensuring technical insights are aligned with business value.
Juan has researched and developed technology on semantic data virtualization, graph data modeling, schema mapping and data integration methodologies. He pioneered technology to construct knowledge graphs from relational databases starting in the mid 2000s, resulting in W3C standards, research awards, patents, software and his startup Capsenta acquired by data.world in 2019. Juan strives to build bridges between academia and industry as former co-chair of the LDBC Property Graph Schema Working Group, member of the LDCB Graph Query Languages task force, standards editor at the World Wide Web Consortium (W3C). Juan continues to be an active member of the scientific community, with 10k+ citations, through academic research partnerships, advising students, and member of data and AI scientific conference committees.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Knowledge graphs and ontologies are finally having their well deserved moment because of AI. But for many of us, this work started way long before the hype.
I started working on knowledge graphs and ontologies (i.e. semantic web) in 2005, developing research-backed approaches to derive knowledge graphs from relational databases. Since then, I’ve taken that work beyond academia, into startups, products, and enterprises struggling to turn data into shared understanding.
This talk shares the scar tissue. This isn’t just a retrospective talk. It’s a practical guide grounded in passion and determination.
I’ll cover the patterns I’ve seen repeat for 20 years: why efforts succeed and fail, why it is still a challenge, why teams underestimate knowledge work and get stuck, how organizations can build semantic foundations without boiling the ocean, architectural approaches…and what I wish I had understood earlier technically, organizationally, and culturally.
If you’re building, buying, or betting on knowledge graphs today, this talk is an honest no-bs set of lessons learned from someone who’s been stubbornly working on the same core problem for 20 years, and is still learning. This talk may save you money and years of pain.
Nadia Rauch
Nadia Rauch
About the Speaker:
Nadia is an AVP of AI Engineering at Chubb, where she leads teams building AI and machine learning solutions at scale. With 14+ years of experience building products and working with
data across multiple roles — including leading Nalys’ Machine Learning and Data Engineering teams — she brings a wealth of expertise to her current position. Nadia holds a Bachelor’s
and Master’s degree in Computer Engineering from the University of Florence and specialises in Knowledge Bases and Ontologies. During her doctoral studies, she developed a pioneering
Smart City system, an innovative project which showcased her ability to harness data from diverse sources to create actionable insights for urban planning and management. Additionally,
Nadia is a dedicated advocate for diversity in tech, serving as a committee member for TMLS – Women x AI, where she empowers young women entering the field.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We built a multi-agent pipeline to automate a complex, document-intensive enterprise workflow. The initial system ran sequentially through a chain of specialized agent roles, each handling a distinct task in the process, using a single frontier model throughout. It worked. It was also slow and expensive, and we didn’t know why.
Rather than optimize blindly, we profiled first. The results were not where we expected: 67% of total latency came from a small minority of the agent roles, and the primary bottleneck was not the model — it was sequential chaining: independent work was being processed one step at a time instead of concurrently. Identifying and parallelizing the roles with no inter-dependency reduced end-to-end runtime by roughly 40%.*
The second experiment compared model tiers (frontier, mid-tier, lightweight) across each agent role independently, measuring accuracy, latency, and cost per role. The finding cuts against the default assumption: the roles that appeared most cognitively demanding required frontier models, but the accuracy gap was smaller than expected. The roles where model downgrade failed were the ones responsible for precise structured extraction — tasks where errors propagate silently downstream and surface only at the output. Swapping to mid-tier models on roles that tolerated it reduced per-run cost by ~40%, with an estimated <2% accuracy impact at the role level and <1% at the pipeline output level.*
The third experiment compared orchestration harnesses — evaluating how the choice of agentic framework affects runtime overhead, observability, and the ease of implementing the parallelization and model-swap changes described above. Framework choice turned out to matter more than expected for operational concerns: debugging multi-agent failures, tracing costs per agent, and modifying execution flow without rewriting pipeline logic.
Operating in a regulated enterprise environment added one constraint worth naming: every agent decision needs to be auditable. This ruled out certain optimization shortcuts that would have been acceptable in other contexts and shaped how we defined “accurate enough” per role.
This talk covers the profiling methodology, the per-role model comparison framework, the parallelization decisions, the harness comparison, and the operational lessons — including what we’d instrument from day one if we rebuilt the system today.
*Cost and accuracy figures for the parallelization and model-tier experiments are preliminary; the formal evaluation is in progress and will be updated with measured results before the talk.
Kumaran Ponnambalam
Kumaran Ponnambalam
About the Speaker:
Kumaran Ponnambalam is a technology leader with 20+ years of experience in Generative AI, Machine Learning, Data and Analytics. His focus is on creating robust, scalable Gen AI models and services to drive effective business solutions. He is currently leading Generative AI initiatives at Cisco, building next-generation AI innovations and products to help enterprises. In his previous roles, he has built conversational bots, ML platforms, data pipelines and cloud services. A frequent speaker at technology conferences, he has also authored several courses on the LinkedIn Learning Platform in Generative AI and Machine Learning.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This session examines how a traditionally static enterprise agent workflow was redesigned around AI-driven adaptation. In most enterprise settings, agent behavior is defined through fixed prompts, rules, and workflows that apply broadly across users and tenants. We redesigned that model into an adaptive agent architecture built around explicit policy layers, feedback loops, and governed personalization. Structurally, this changed the system from one-time configuration to continuous learning: policies were separated from model reasoning, policy scope was organized across global, tenant, and user levels, and agent behavior was connected to both explicit and implicit feedback signals that could influence future decisions.
The session will focus on what changed in the system architecture and operating model to support this shift. It will cover how policy-driven personalization was introduced, how feedback became part of the runtime and improvement loop, and how adaptation was managed without losing enterprise control. It will also discuss the practical challenges that emerged when moving from static agents to adaptive ones, including handling conflicting user and tenant needs, keeping personalization aligned with core guardrails, and making behavior changes observable and manageable in production.
Attendees will walk away with a practical understanding of how to redesign enterprise agent workflows around adaptive AI capabilities. They will learn the core concepts behind Personalized Experiential Learning, the structural changes needed to support policy-based adaptation, and the key lessons and best practices for making adaptive agents workable in real enterprise environments. They should leave better equipped to think about how to introduce governed personalization, feedback-driven improvement, and policy evolution into their own AI agent systems.
Tony Blank
Tony Blank
About the Speaker:
Tony Blank is a Staff AI Engineer at Trust & Will, where he leads the internal AI tooling discipline — equipping engineers with background agents, harnesses, and infrastructure that turn AI into measurable engineering velocity. He brings 15 years of developer-community leadership at Twilio, SendGrid, and Agora to the work, bridging deep technical work with business outcomes for developer-first businesses. Tony speaks on the patterns that survive contact with production agents: API standards, eval discipline, build-vs-buy economics, and the org changes as “AI Engineer” splits into specialized roles. Previously seen at APIDays NYC. Based in Denver.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We stopped equipping humans and started automating their worst hour. Instead of teaching people to fish, we built the pond: precise background agents pointed at repeated, skill-independent time-sinks where ROI was measurable on day one. The first build was a bug-triage agent wired to GitHub, Jira, and Notion; every agent that followed was measured the same way — Time Saved Per Task (TSPT) per week — against an amortized cost model (build + inference + maintenance vs. hours saved × loaded hourly cost).
Structurally, the change was moving from a portfolio of per-person tool subscriptions to a portfolio of scoped, measurable background agents with a pre-build attribution conversation baked in. The decision rubric — Sharpness × Cadence × Skill Tax → Pond / Boat / Net / Wade — tells you when to automate fully, when to augment, when to keep a human in the loop, and when the math doesn’t work yet.
Attendees will leave able to: score their own candidate workflows on that rubric, instrument TSPT, build a defensible amortization story, and avoid the most expensive failure mode in AI enablement — “give everyone the tools” — by building the pond first.
Mahtab Dhaliwal
Mahtab Dhaliwal
About the Speaker:
Mahtab Dhaliwal is an ML Engineer at Shopify, building agentic pipelines for commerce intelligence and fraud detection. Her work spans prompt optimization, production infrastructure, and evaluation — focusing on turning experimental AI flows into reliable, scalable systems that support real-world decisions. She is keen on learning the latest and applying it solve problems, building systems that anticipate tomorrow’s needs.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Shopify merchants can customize almost every aspect of their storefronts. That flexibility makes even basic questions about a shop’s brands, shipping policies, or return terms difficult to answer consistently at scale. Our first approach sent selected pages to a frontier model in a single prompt. It worked initially, but missing evidence, competing objectives, and inference cost limited its quality and coverage.
This talk traces the redesign of Shop Lens from one-shot extraction to a DSPy ReAct agent, then to three specialized agents for shop profiling, fraud signals, and merchant category codes. I’ll explain how we scoped each agent’s context and tools, reused fetched pages through shared infrastructure, and optimized each agent independently with GEPA.
Because live storefronts change, we built ShopSnap to capture and replay stores at annotation time, enabling reproducible evaluation and separate optimization and held-out datasets. We also separated batch orchestration, CPU-based agent execution, and self-hosted OSS model inference across Flink, Kubernetes, and GPU infrastructure.
In the results shared in my earlier talk, the redesigned system achieved approximately 75× lower inference cost and close to 2× improvement on selected quality metrics compared with the original one-shot approach. Attendees will leave with concrete patterns for task decomposition, evidence access, replay-based evaluation, and model selection—and an understanding of how those architectural choices compound.
Deji Andrew
Deji Andrew
About the Speaker:
Deji Andrew is Manager of Systems & Data Platforms at Niagara Bottling, a $3B+ beverage manufacturer operating 50+ production facilities. He works on ML and data infrastructure supporting supply chain, warehouse operations, and production planning. Over the past several years he has built and deployed production ML systems across the company’s operations – including a warehouse pallet retrieval predictor, a real-time production overrun correction pipeline, and a downtime prediction system running across nine facilities – alongside the data infrastructure that supports them. His work focuses on the specific failure modes that emerge when ML systems act on physical control systems rather than digital ones. He holds a BS in Engineering from Harvey Mudd College and is completing an MS in Data Science at the University of Pennsylvania.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
A LightGBM model was deployed across production lines to predict and correct overruns mid-job, reducing average overproduction from +24 cases to +2 on historical data. In pilot, the corrections worked: overruns vanished. But a closer look at the production data showed something different. When the order target was reduced by 21 cases, the line overproduced the new target by 18. The system was not being corrected – it was negotiating with the correction, regenerating overrun relative to whatever target the machine was now working toward. Compounding this: corrected jobs left no trace in the training data that a human had intervened, meaning the model, on retraining, would learn to stop flagging the problem it had been built to solve. This talk walks through the pilot data, explains why conservatism in the loss function wasn’t enough, and argues that the real gap isn’t in the model – it’s in the experiment infrastructure that doesn’t exist between a prediction and a PLC. Attendees will leave with a concrete mental model of reflexivity and contamination in industrial ML deployment, a framework for when not to close the automation loop, and specific reasons the standard MLOps playbook (holdouts, impression logging, A/B frameworks) doesn’t transfer to systems that act on physical state.
Maitrik Patel
Maitrik Patel
About the Speaker:
Maitrik Patel is a Senior Engineering Manager, leading ML and AI platform engineering at consumer scale. He has built production infrastructure spanning ML training pipelines, AI inference systems, and agent workflow orchestration for a platform serving hundreds of millions of users. He spent a decade as a web developer, UI engineer, and design technologist at Walt Disney, Sony, and DocuSign – building enterprise design systems and front-end platforms. Maitrik is a co-author of ASTRA-bench (arxiv.org/abs/2603.01357), a published benchmark evaluating tool-use agent reasoning with personal user context. He is an IEEE Senior Member, Forbes Technology Council member, and a reviewer for CVPR, ACL, and ICLR 2026 workshops.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Running ML training, AI inference, and agent workflows on three separate orchestration systems is an operational debt that compounds with every new workload type – one system per abstraction means three on-call rotations, three observability stacks, and three onboarding paths. We collapsed this to a single execution layer by designing a type-safe task interface that accommodates non-deterministic agent execution alongside deterministic ML pipelines, backed by container-native isolation and a workload-aware shared scheduler. The unification exposed a hard failure class that static ML infrastructure never encounters: agent tasks are not idempotent, and ML-derived retry semantics caused downstream state corruption in early production until we rebuilt retry logic around explicit checkpointing and action journals. Attendees leave with a four-layer architectural blueprint for unifying these workload types, an honest account of the scheduling contention and retry failures encountered in production, and the specific interface design decisions that made incremental migration rather than forced rewrite the path to adoption.
Kalpesh Sutaria
Kalpesh Sutaria
About the Speaker:
Kalpesh Sutaria is Director of Engineering at NVIDIA, where he leads engineering for the NeMo/Nemotron Retriever inference stack and the MLOps platform behind it — the serving layer that productionizes NVIDIA’s retrieval models for enterprise deployment.
Across 20 years at NVIDIA and AWS he has built AI/ML platforms, launched Amazon Lex from an internal 0-to-1 bet to GA, and run enterprise and payments systems at scale.
Kalpesh holds six granted U.S. patents and cares most about the seam between research and production, where performance, reliability, and cost decide whether great models ever reach users.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
A leaderboard win is not a product. Nemotron Retriever’s models rank #1 on RTEB — but topping a benchmark and running efficiently inside a customer’s production are two very different problems. This talk is the engineering story of closing that gap.
I’ll walk through our journey rebuilding the Retriever inference stack in Rust with a single obsession: treating state-of-the-art performance as a first-class product requirement, not a post-hoc optimization. We’ll get concrete about the decisions that mattered — what we measured, where we spent effort, the throughput, memory, and footprint wins we chased, and how dramatic efficiency gains unlocked deployment scenarios (including edge and on-device) that simply weren’t possible before. I’ll also share what didn’t work, the tension between shipping fast and building durable, and how we leaned on AI-assisted development to move faster than the roadmap assumed.
You’ll leave with a mental model for treating inference performance as a product and go-to-market lever — and a practical playbook for taking research-grade models to production without leaving speed, cost, or reach on the table.
Key takeaways:
- Why “performance is the product”: efficiency is an adoption and go-to-market lever, not a cost center.
- Concrete decisions from a Rust-based inference rebuild — what to measure, where to optimize, and how to know it worked.
- How latency and footprint reductions open entirely new deployment surfaces (edge, on-device) for the same models.
- Navigating the research-to-production seam: turning SoTA benchmark results into reliable, cheap, fast serving customers can bet on.
Who should attend / level? AI engineers, ML platform/infrastructure engineers, and engineering leaders who build, deploy, or operate inference and serving systems. Intermediate.
Anupama Garani
Anupama Garani
About the Speaker:
Anupama Garani is VP, AI Engineering at PIMCO, where she has led production GenAI since 2023. She leads architecture, evals, and products that help sales and tech teams work from trusted company knowledge. Before PIMCO she was a BI Analyst at Goldman Sachs Ayco. She has spoken at 20+ AI workshops, conferences, events, and university meetups. She moderated and MC’d sessions at MLOps World | GenAI Summit 2025, and led workshops at UT Austin McCombs and Women in Data Science (WiDS). She serves as a program committee member and reviewer for AAAI/ACM AIES and ACL TrustNLP, founded the Central Texas ODSC chapter, and runs TRUST AI, her 8-week AI program for high school students across the US.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
RAG systems fail in many ways. One metric at one stage cannot find them. This talk gives a full map: 33 failure modes across 7 pipeline stages: ingestion, representation, retrieval, generation, evaluation, deployment, and agentic orchestration. The map comes from a structured review of 48 sources, with a formal definition, a visible sign, and an evidence grade (Strong / Moderate / Limited) for each mode. It was published at TrustNLP 2026 (ACL). One finding matters most for agents in production: research studies retrieval and generation well, but it studies representation, evaluation, and agentic orchestration far less. 12 modes have no dedicated peer-reviewed empirical evidence, and all 8 agentic modes are in that set. You will learn to find each failure by stage, test for it before users do, and fix the stage that broke, not only the final answer.
Vashishtha Patil
Vashishtha Patil
About the Speaker:
Vashishtha Patil is a Senior Applied Scientist at Amazon, where he leads science research on efficiency for Alexa+ Smart Home. Across 13 years at Amazon and Qualcomm he has designed and optimized AI/ML algorithms for latency-constrained production systems, work reflected in multiple granted U.S. patents. He serves as a peer reviewer for top-tier AI conferences and is a frequent judge at competitive AI events and startup pitch competitions
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Autonomous research AI agents are becoming remarkably capable, but running long investigations can be expensive—and more computation does not always produce better results. How can we make autonomous research both affordable and productive?
This talk explores an approach that combines smaller AI models, selective frontier guidance, and resource-aware research orchestration. Drawing on experiments and a recorded demonstration, we’ll examine where these systems spend their effort, what helps them make meaningful progress, and when additional model capability earns its cost.
Attendees will gain a framework for evaluating the trade-offs between capability, cost, and research quality, along with lessons for building autonomous workflows that use limited resources effectively. We’ll close with the open challenges that stand between promising demonstrations and scalable autonomous research.
Poonam Lamba
Poonam Lamba
About the Speaker:
Poonam Lamba is an AI Infrastructure leader specializing in scaling Reinforcement Learning (RL) and Inference workloads, and high-performance computing environments. With deep expertise in GPU/TPU optimization, orchestration, and workload efficiency, Poonam works at the forefront of AI platform engineering to enable next-generation model training and inference.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We redesigned distributed GPU orchestration for RL post-training and batch inference in the open-source llm-d platform. Structurally, we replaced static GPU/TPU locking with a three-tier co-operative time-slicing system:
- Application Layer: Workloads signal phase boundaries (rollouts, training, batch inference) via explicit acquire() and yield() APIs.
- Cluster Orchestrator: Manages lock queues to dynamically interleave complementary jobs onto shared hardware during idle phases.
- Node Snapshot Agent: Executes fast sub-second state swaps between GPU/TPU VRAM and host DRAM, enabling instant context switching without container restarts.
Attendees will walk away with:
- Drive 70%+ GPU/TPU Utilization: Understand how time-slicing reclaims idle hardware during RL loops and batch inference without impacting convergence.
- Architect Rapid Memory Swapping: Apply VRAM-to-DRAM snapshotting strategies for ultra-fast GPU/TPU context switching.
- Deploy on Kubernetes: Configure llm-d and K8s orchestrators to interleave RL and batch inference on shared clusters.
Zachary Hamilton
Zachary Hamilton
About the Speaker:
Zach is a Braintrust Solutions Engineer with 10 years of experience in data infrastructure and observability.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Building agentic software requires more than adding an LLM to the traditional software development lifecycle. When behavior becomes non-deterministic, teams need a new feedback loop connecting what happens in production to how they evaluate, debug, and improve their systems. This talk will show how to build that loop: identifying real failure modes from production traces and human feedback, turning them into reproducible datasets and evaluations, defining success using both technical and business outcomes, and using automated and human judges to measure whether changes actually improve the system. We’ll also explore where automated evaluation breaks down, how to calibrate LLM-as-a-judge, and how continuous evaluation helps teams detect regressions and emerging behavior. Attendees will leave with a practical framework for moving from production > failure discovery > datasets > evals > iteration > production, turning the development of agentic systems into a measurable engineering discipline.
Naresh Kumar Batthula
Naresh Kumar Batthula
About the Speaker:
Principal Developer Advocate on DraftKings’ AI Platform. I bridge engineering and leadership by translating AI strategy into platform capabilities engineers actually adopt, while carrying the developer voice back into the roadmap.
After 15+ years leading engineering at global scale, I made a deliberate move from running engineering orgs to going deep on the AI platform problem because that is where the leverage is right now. My focus is making AI a dependable, governed part of how teams build software, not a pile of demos.
What I focus on:
- Agentic development and AI-DLC: I created the AI-Driven Development Life Cycle, a multi-agent framework spanning inception, construction, and quality gates, with an L0–L4 maturity model that gives teams a clear path from experimentation to fluency.
- Productionizing AI agents: Helping take internal agents from prototype to production with the guardrails, boundaries, and hardening needed to run them safely, including DraftCode for AI code review, autonomous PR remediation, and skill-based auto-approval, and HelpAI for agentic troubleshooting.
- Evaluation and trust: Building LLM evaluation frameworks for Code review tooling, RAG based knowledge systems, covering retrieval quality, grounding, and faithfulness through LLM-as-judge workflows, CI release gates, and production observability, so quality and access-control regressions are caught before they reach users. My north star: reward evidence, not confidence.
- Platform and roadmap influence: Helping shape the AI Platform’s technical direction across multi-LLM access, MCP-based tooling, enterprise knowledge management, governance, RBAC, and cost-control standards.
- Adoption and enablement: Building the community support, documentation, and developer experience that turn platform capability into real, measured adoption — the part of DevRel that actually moves the needle.
I collaborate with LLM vendors including Google, OpenAI, and Anthropic, and evaluate platforms across Gemini Enterprise, OpenAI, and AWS Bedrock. My work spans the modern GenAI stack: LLM gateway, Agentic Identity, Governance, network boundaries, Remote agents, LLM Obs using Datadog and MlFlow, N8N, MCP’s, Agentic Memory, A2A, and Agent development frameworks.
Before DraftKings, I led engineering for Walmart Global Tech’s next-generation cloud checkout platform across five international markets: the US, Canada, Japan, the UK, and Mexico.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
AI has speed up code generation, but the rest of the software development lifecycle has not necessarily kept pace. At DraftKings, as we rolled out AI tooling across our engineering org, we saw pressure move to PR review, testing, integration, and release confidence.
This talk shares where those bottlenecks emerged and takes a closer look at DraftCode, our code review agent. Its first version took a diff and returned generic feedback. That wasn’t useful enough. We then worked with the teams responsible for compliance, security, and platform standards to make their rules available to the agent at the global, tool, team, and service levels. I’ll walk through one concrete example, a security rule that flags hardcoded secrets, show how DraftCode applies it, and how reviewers went from catching these issues by hand to having them flagged automatically.
We also tried to build a complete AI-assisted SDLC loop. That approach saw limited adoption because many teams already had tools in place. That experience changed where we put our platform effort: on problems individual engineering teams could not readily solve themselves.
Fuzail Khan
Fuzail Khan
About the Speaker:
Senior machine learning engineer at Meta working on AI training infrastructure and performance optimization. At Meta, I’ve been primarily working on recommendation systems and optimizing and scaling these systems to over 4 billion users across Meta’s apps surface.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
The ranking and recommendation systems landscape is being transformed in the generative era. This talk reports on the experience of building and shipping the training infrastructure behind the first generative recommender system at Meta, covering both stages of the recipe: pre-training to acquire the generative capability and post-training RL to align generation with the ranking objective.
Generative recommenders bring distinct challenges to end-to-end performance and scalability. This primarily arises from a mixed architecture that consists of both recommender-native large sparse embedding tables and LLM-based decoders. This then generates semantic IDs translating to real-world use cases such as finding the right advertisement for a given user. We inherit the communication profile of a sparse recommender as well as the autoregressive nature of a large language model for which an end-to-end systems blueprint simply does not exist.
For pre-training, we make the significant change from discriminative to generative recommendation. We talk about feature processing, data loading and the user-modeling path while building the LLM decoder and semantic ID tokenization on top while addressing real-world challenges and end-to-end LLM performance analysis.
Reinforcement learning was implemented as an extensible framework where rewards, losses, RL algorithms, reference models and generation strategies are pluggable. We’ll walk through the reward design and the end-to-end post-training pipeline that prioritizes scalability at the end-to-end scale at Meta.
We’ll close with end-to-end optimization work – redundant computation elimination, specialized kernels for very short sequences, dense-sparse pipelining and hybrid embedding placement that satisfies large-scale performance requirements in compute and throughput.
We believe the key output from this session is leaving attendees with a crisp understanding of generative recommendation systems at large-scale. We want to ensure we stay practical and address real-world challenges that are relevant to developers and builders in AI infrastructure today. We do this by diving into the technicalities of enabling a large-scale generative retrieval system as well as concrete performance optimizations that enable the shift from a research prototype to a highly-optimized performant system. We also feel the training systems and infrastructure-side of LLMs is rarely addressed at least relative to model architecture and quality, and we hope to fill that gap with this session.
Balaji Varadarajan
Balaji Varadarajan
About the Speaker:
Balaji Varadarajan is a Staff Engineer at DigitalOcean, where he is a founding engineer of the company’s dedicated LLM Inference platform, architecting infrastructure that serves billions of tokens daily. With 14+ years in scalable systems and 5+ years in production AI/ML, he previously led ML engineering at Target, building recommendation and RAG systems serving millions of guests, and held senior roles at HubSpot and CCC Intelligent Solutions. He holds a pending LLM patent, is an IEEE Senior Member, and is pursuing an MS in AI at UT Austin.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We will start with two diagnostic tools: Little’s Law which translates your QPS and target latency into the concurrency and batch size your system actually needs and the roofline model which tells you whether prefill (compute-bound) or decode (memory-bound) is your real constraint before you reach for a fix..
Will walk through the levers that move the needle most – MoE vs. dense architecture and what that does to your communication pattern FP8/INT4 quantization and where it costs you accuracy, TP/EP/DP parallelism and when each is the right tool and batching to the B_sat knee before compute-bound latency hockey-sticks..
Will cover concrete tuning playbooks for two workload types:
- latency-sensitive traffic like chat and agents where you’re optimizing TTFT and P99 ITL with chunked prefill and modest batch sizes
- Throughput-sensitive traffic like batch summarization and evals where you push past B_sat and let queuing work in your favor.
Most production systems live in both worlds at once, so we’ll need a balanced configuration to tackle both.
You will leave with a 7-step decision framework – characterize your workload, pick your model and quantization, benchmark on candidate hardware, find the knee, size your deployment with Little’s Law, calculate total cost of ownership, and plan for autoscaling that you can apply to your own inference stack..
Zachary Carrico
Zachary Carrico
About the Speaker:
Zac is a Senior AI/ML Engineer at Apella, specializing in AI products for improving surgical operations. He has a deep interest in healthcare applications of AI, and has worked on cancer and Alzheimer’s disease diagnostics. He has end-to-end experience developing AI systems: from early research to serving thousands of daily customers. Zac is an active member of the Data and AI community, having presented at conferences such as Ray Summit, TWIML AI, Data Day, and MLOps & GenAI World. He has also published eight journal articles. His passion lies in advancing AI and streamlining the deployment and monitoring of models, reducing complexity and time. Outside of work, Zac enjoys spending time with his family in Austin and traveling the world in search of the best surfing spots.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
An on-call agent swarm multiplies the cost of production alert noise. They happily burn tokens investigating alerts and coming up with plausible sounding, but incorrect root causes. To mitigate this requires understanding their biases and an alert triage layer in front of the agent swarm.
Per alert, that layer answers three questions: is this worth investigating, does it belong with other alerts as one incident, and how many agents does it get? It feeds co-firing alerts into the swarms context and throttles agent parallelization.
We evaluated the ability of agents to reach the same root cause humans achieved on incident post-mortems. We quickly found that there’s a lot of context that agents didn’t have access to or didn’t consider investigating. Understanding the context that is most helpful to identifying a root cause also allowed us to better guide agent task ownership. Through iterative refinements we improved their performance.
Identifying the production alerts that provide the greatest signal-to-noise allowed us to deploy the swarm with precision. Having humans review the swarm’s conclusions provides new training data to optimize their behavior with.
You’ll leave with a playbook for deploying agents for on-call alert triaging, mistakes we made that we hope you won’t repeat, and an orchestration layer that keeps agent cost bounded.
Ravi Satya Durga Prasad Yenugula
Ravi Satya Durga Prasad Yenugula
About the Speaker:
Ravi Satya Durga Prasad Yenugula is a Senior/Lead Data Engineer at 2K Games and an AI infrastructure researcher focusing on efficient LLM deployment and open-source data systems. His work spans large-scale enterprise data architecture, open-source contributions to core projects like Pandas and Apache Superset, and research on model downscaling and knowledge transfer (including ScaleOp). Ravi bridges the gap between deep learning research—such as model downscaling, quantization, and structural pruning—and scalable, production-grade MLOps pipelines.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
1. Structure-respecting selection beats dense blending by 15-175× at zero-shot. 2. Compensation recovers function-level quality but has an ill-conditioning boundary (5× donor scale); rescale complements it. 3. The norm-placement rule is architecture-agnostic — it should transfer to RMSNorm, SwiGLU, GQA. 4. Distillation interacts asymmetrically: helps compensated init (88→83 ppl), hurts pruning init (92→97 ppl). 5. At matched budget, initialization beats fine-tuning; at large budgets, they converge
Joe Reis
Joe Reis
About the Speaker:
Joe Reis is the best-selling co-author of Fundamentals of Data Engineering, author of Mixed Model Arts, and founder of Practical Data Media. With 25+ years in the data industry, he helps audiences build durable mental models for AI, data engineering, architecture, and modeling.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
For decades, data engineering has focused on the data engineering lifecycle: building pipelines, orchestrating workflows, managing storage and compute, and overseeing the undercurrents – security, architecture, data governance, and more. But as AI transitions to systems that take action, infrastructure is no longer the main bottleneck. Meaning and clarity in data are.
When machines take actions, semantic ambiguity becomes catastrophic. This talk explores the shift from machinery to meaning and why the discipline’s future depends on a renaissance in data modeling. We will examine why rigid dogmatism and unstructured chaos both fail modern systems, and how adopting Mixed Model Arts – a pragmatic, multi-paradigm approach to modeling context, semantics, and relationships – provides the foundation machines need to understand our world and act responsibly within it.
Eric Riddoch
Eric Riddoch
About the Speaker:
Eric has worked in MLOps for most of his career. He led the ML Platform efforts at BENlabs, then Pattern (largest seller on Amazon.com), and then Subquadratic (a frontier LLM research lab working on cheaper attention mechanisms). Eric authored two best-selling MLOps courses through MLOps Club, and currently works as a forward deployed engineer at Flexion.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
There are two main obstacles that make LLMOps harder than classical MLOps: data gravity (the models and datasets are big) and GPUs. How do I find ‘affordable’ GPUs? Neoclouds are cheaper, but they are of mixed quality. What about minimizing GPU idle time? You can treat your GPUs like a job queue, but how do you share between training and serving workloads? How do you make a queue ‘ergonomic’ for researchers to use? How do you make runs reproducible and avoid dependency hell? uv is great, but it makes you ‘solve’ dependencies by hand when it comes to CUDA. I do not have all the answers, but I will share many things I learned through working on these problems.
Siddharth Jain
Siddharth Jain
About the Speaker:
Siddharth (Sid) Jain is an AI engineering manager at OpenAI, where he leads work on enterprise AI systems and governed agent workflows. He has more than 10 years of experience across AI, finance, data, and enterprise platforms, with a focus on safe automation, retrieval, auditability, and production controls. He speaks and writes about turning AI prototypes into reliable systems that organizations can trust.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Teams often give an agent a service credential, add a human approval step, and call the workflow governed. That design breaks down when the agent can revise payloads, retry writes, chain tools, or act across systems with different permission models. The result is an accountability gap: the organization can see that a service account acted, but not necessarily who authorized the business intent, which payload was approved, whether a retry duplicated work, or what changed between proposal and execution.
This session presents a production control model for agent workflows that make consequential writes. It shows how to separate the business-intent identity from the concrete operation; classify tools by impact; issue short-lived, least-privilege credentials only after validation; bind human approval to a canonical payload hash and policy version; execute through controlled services with idempotency keys; and reconcile external state before declaring success. It also covers the evidence record needed to answer four operational questions: who requested the action, what the agent proposed, what policy and human approved, and what the downstream system actually did.
The talk focuses on failure modes that appear after the demo works: stale approvals, overbroad agent permissions, duplicate writes after retries, silent policy changes, and actions whose outcome is unknown. Attendees will leave with a lifecycle they can map to their own platform, concrete interfaces between model output and deterministic controls, and an audit schema that supports incident response, compliance review, and day-to-day operations without turning every agent into a bespoke security project.
What You’ll Learn:
- Do not let the model hold broad standing credentials; mint scoped identity for each authorized operation.
- Bind approval to material fields, a canonical payload hash, and the policy version so edits invalidate stale consent.
- Treat every consequential write as a durable state machine with explicit recovery paths, not as a single tool call.
- Separate retry from reconciliation so an ambiguous timeout does not become a duplicate action.
- Design the audit record as an operating interface: it should reconstruct who requested, what was proposed, what was approved, what executed, and what the downstream system recorded.
Prerequisite Knowledge:
Dr. Amir Feizpour
Dr. Amir Feizpour
About the Speaker:
Dr. Amir Feizpour is Founder, CEO and Chief Scientist of Aggregate Intellect, building a generative business brain for service and science companies, and serves as a fractional Chief AI Officer for several startups. Previously he led NLP product at Royal Bank of Canada and did quantum computing research at the University of Oxford, resulting in publications and patents. He holds a PhD in Physics from the University of Toronto. Amir has built a global community of 5,000+ AI practitioners, advises at MaRS Discovery District, and leads R&D collaborations with McGill, the University of Toronto and Amii.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
An evidence-driven workflow for building production AI agents: testing assumptions with prompts and real users before writing code, mapping context flow before setting agent boundaries, and letting decomposition emerge from evidence rather than imposed architecture. Attendees will leave with a reusable ideation workflow, a setup guide for AI development assistants, and a framework for evidence-based architectural decisions, illustrated through a production case study of an agentic meal-planning system built in a few weeks.
Michael Perry
Michael Perry
About the Speaker:
Software is math. Every class is a theorem. The compiler is the proof. And unit tests check our work. Michael L Perry has built upon the works of mathematicians like Marc Shapiro, Pat Helland, and Leslie Lamport to develop a mathematical system for software development. He is the author of The Art of Immutable Architecture and the Jinaga Immutable runtime. He applies his love of software mathematics at Improving, a trusted partner for AI-assisted application development.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
AI-assisted development tools can now quickly produce working software that once took months to hand craft. But that software often is based on the average of opinions on which the model was trained. It lacks the precision and expertise of the team that maintains it, the specific stack that they selected, and the particular problem domain that it serves. To produce high quality solutions at scale, you need to build a custom software factory. The factory is built from standard features available in almost all AI-assisted development tools:
- Skills
- Commands
- Agents
- Workflows
Your team constructs it to capture the specific architectural decisions and implementation patterns that they have developed for the project. They maintain the factory over time as they use it to build artifacts like documentation, code, and tests. Every project is different. Don’t install a generic template. Learn how to construct a software factory for your current project to accelerate development while improving quality and consistency.
Upal Saha
Upal Saha
About the Speaker:
Upal Saha is the co-founder and CTO of bem, infrastructure for building AI-native products that run critical business processes reliably, at scale. He leads the engineering team that runs AI over millions of documents, images and videos a month for customers in logistics, fleet management, automotive and financial services. Before bem he was an engineering manager at Silo and a technical lead at Yext, and spent time on the investing side at Akido Labs. Based in San Francisco.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This is a build-and-break lab for engineers who already know that a demo is not a system. In the first fifteen minutes every attendee stands up a working extraction pipeline against a real invoice, with no schema authoring, and gets structured JSON back. Then we spend an hour breaking it the way production does, and fixing each break with a pattern that transfers to any stack.
Break one: the wrong document. We feed a bill of lading glued to an invoice into the invoice pipeline and watch it confidently produce garbage. Fix: classify before you extract, and make the graph deterministic even though every step inside it is a model. Break two: the answer that is probably right. Models do not tell you how confident they are, so we compute per-field confidence, find the fields that fall below 95%, and route those, and only those, to a human whose correction feeds back into the system. Break three: the messy string. “10 cases organic gala apples, 88 ct” has to become one SKU; we show why canonicalization is its own step, how to score matches, and where to set the threshold for review. Break four: the hostile input. Everyone runs an image whose pixels read “ignore all previous instructions” and we discuss, with the result on screen, what it means to treat inbound data as data rather than control. We close by labeling a handful of outputs and running a regression test between two versions of the pipeline, because evaluation is a loop, not a phase.
Attendees leave with a running pipeline in their own account and five patterns they can apply on Monday regardless of vendor: classify-then-extract, confidence scoring for models that lack it, threshold-based exception routing with feedback, semantic security checks on inbound data, and versioned regression testing. The platform used in the room is ours; the lessons are not.
Matt Foley
Matt Foley
About the Speaker:
Matt partners with enterprise clients across sales, onboarding, and technical support, helping teams build reliable, AI powered document workflows. His work sits at the intersection of customer strategy and hands on technical problem solving.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This is a build-and-break lab for engineers who already know that a demo is not a system. In the first fifteen minutes every attendee stands up a working extraction pipeline against a real invoice, with no schema authoring, and gets structured JSON back. Then we spend an hour breaking it the way production does, and fixing each break with a pattern that transfers to any stack.
Break one: the wrong document. We feed a bill of lading glued to an invoice into the invoice pipeline and watch it confidently produce garbage. Fix: classify before you extract, and make the graph deterministic even though every step inside it is a model. Break two: the answer that is probably right. Models do not tell you how confident they are, so we compute per-field confidence, find the fields that fall below 95%, and route those, and only those, to a human whose correction feeds back into the system. Break three: the messy string. “10 cases organic gala apples, 88 ct” has to become one SKU; we show why canonicalization is its own step, how to score matches, and where to set the threshold for review. Break four: the hostile input. Everyone runs an image whose pixels read “ignore all previous instructions” and we discuss, with the result on screen, what it means to treat inbound data as data rather than control. We close by labeling a handful of outputs and running a regression test between two versions of the pipeline, because evaluation is a loop, not a phase.
Attendees leave with a running pipeline in their own account and five patterns they can apply on Monday regardless of vendor: classify-then-extract, confidence scoring for models that lack it, threshold-based exception routing with feedback, semantic security checks on inbound data, and versioned regression testing. The platform used in the room is ours; the lessons are not.
Rajiv Shah
Rajiv Shah
About the Speaker:
Rajiv Shah is the Agentic AI Engineer at OpenHands with a passion and expertise in Practical AI. He focuses on enabling enterprise teams to succeed with AI. Rajiv has worked on GTM teams at leading AI companies, including Hugging Face in open-source AI, Contextual AI in context engineering, Snorkel in data-centric AI, Snowflake in cloud computing, and DataRobot in AutoML. He started his career in data science at State Farm and Caterpillar.
Rajiv is a widely recognized speaker on AI, published over 20 research papers, been cited over 1000 times, and received over 20 patents. His recent work in AI covers topics such as sports analytics, deep learning, and interpretability.
Rajiv holds a PhD in Communications and a Juris Doctor from the University of Illinois at Urbana Champaign. While earning his degrees, he received a fellowship in Digital Government from the John F. Kennedy School of Government at Harvard University. He is well known on social media with his short videos, @rajistics, with over 100k followers.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
You can start a simple agent with a model and a prompt. But you soon realize that improving it requires adjusting what the model can see, what it can do, how it remembers progress, which model handles each step, and what evidence allows the work to stop. Harness engineering is how those pieces become one working system.
This workshop uses coding agents to make that system concrete. Participants will work through six decisions: choosing a harness, designing tools and retrieval, placing context and memory, routing models and reasoning, controlling goals and validation, and deciding when a task benefits from multiple agents. Through hands-on experiments, we will change these settings and inspect what happens in the trace and final result.
Each exercise is paired with an overview of current research on agent harnesses. Together, the research and experiments show how the major components interact and where each approach reaches its limits. Participants will leave with a practical way to reason about the whole agent, not just its model.
Chris Alexiuk
Chris Alexiuk
About the Speaker:
Chris Alexiuk is a deep learning developer advocate at NVIDIA, working on creating technical assets that help developers use the incredible suite of AI tools available at NVIDIA. Chris comes from a machine learning and data science background, and he is obsessed with everything and anything about large language models.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
In this workshop we’ll stand up NemoClaw end to end: install the reference stack, get OpenClaw running inside the OpenShell sandbox, configure inference routing, and lock down a network policy that survives a multi-hour agent session. We’ll walk through the blueprint, the CLI, and the approval flow, then run a real long-lived agent against it and break things on purpose so you know what the layers actually catch.
Greg Loughnane
Greg Loughnane
About the Speaker:
Dr. Greg” Loughnane is the Founder & CEO of AI Makerspace, where he builds training programs for engineering and data science teams at mid-market and enterprise companies. Since 2021, he has built and led industry-leading Machine Learning education programs. Previously, he served as an AI product manager, university professor, consultant, advisor, and researcher. He loves trail running, wrenching on cars, and is based in Columbus, Ohio.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
In this workshop, we’ll explore best practices for maximizing ROI (and excitement!) for your retrieval system throughout your entire AI product lifecycle. You’ll understand how to make data and retriever decisions at the MVP, PoC, and Production phases of your project.
We’ll center our workshop around the story of Nadia, a second-line support engineer on the service desk of a regional broadband and mobile operator who is looking to build a retriever for her knowledge base. What should she choose based on the shape of the corpus today, and what kind of corpus change should make her reconsider?
Guided by our heroine, we will explore the key aspects of retrieval; that is, getting the best of both semantic and keyword-based search, using agent loops versus RAG/search pipelines, constructing knowledge bases using embedded chunks versus an LLM Wiki approach, and of course, what happens when Nadia expands the scope of the problem to include multimodal image retrieval.
Scott Askinosie
Scott Askinosie
About the Speaker:
Scott Askinosie, PhD, is a Senior Principal AI Engineer at SteelEngine, where he architects the knowledge and memory layer for governed agentic workflows in regulated enterprises. He built SteelEngine’s three-tier governed memory system (session, team, org) with deterministic policy gates, novelty scoring, provenance, and human endorsement checkpoints. Previously he spent six years as Lead Data Scientist and Professor at WGU, building student retention analytics across 487K+ records, and served as Lead Technical Trainer at Weaviate, delivering 20+ enterprise workshops including NATO, FactSet, and Fannie Mae. Scott holds a PhD in Quantitative Biology from the University of Missouri, leads the 400+ member LangChain Austin community, and speaks at ODSC and Reuters Momentum.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
This is a hands-on build. Attendees start with an ungoverned append-only memory on a vector database, break it deliberately, and then rebuild it as a governed memory with tiered scope.
We work through the write path step by step: capturing a candidate memory, scoring it for novelty against what is already stored, attaching provenance, and passing it through a deterministic, fail-closed gate before it is promoted from session scope to team scope to organization scope. Attendees implement the gate themselves rather than calling a library, because the point of the exercise is understanding exactly which decisions must stay deterministic and which can be model-scored.
The second half is adversarial. Attendees inject false facts into their own memory and measure what survives promotion, then tighten their gates until the poisoning is quarantined. We finish with the evaluation question most teams skip: how to tell “”the agent never received the memory”” apart from “”the agent received it and ignored it,”” and why an eval suite that only scores output quality cannot detect a memory layer that has silently stopped working.
Attendees leave with working code, a reference architecture for tiered governed memory, and a test harness for poisoning resistance they can point at their own stack.
Kshetrajna Raghavan
Kshetrajna Raghavan
About the Speaker:
Kshetrajna Raghavan is a Principal ML Engineer at Shopify who builds AI systems from initial experiments through evaluation and deployment. His work spans multimodal product understanding, commerce intelligence, retrieval, and agents, with a focus on turning messy, unstructured information into useful products. He works across models, agent architectures, evaluation, and inference infrastructure to improve quality and reduce cost at scale.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Shopify merchants can customize almost every aspect of their storefronts. That flexibility makes even basic questions about a shop’s brands, shipping policies, or return terms difficult to answer consistently at scale. Our first approach sent selected pages to a frontier model in a single prompt. It worked initially, but missing evidence, competing objectives, and inference cost limited its quality and coverage.
This talk traces the redesign of Shop Lens from one-shot extraction to a DSPy ReAct agent, then to three specialized agents for shop profiling, fraud signals, and merchant category codes. I’ll explain how we scoped each agent’s context and tools, reused fetched pages through shared infrastructure, and optimized each agent independently with GEPA.
Because live storefronts change, we built ShopSnap to capture and replay stores at annotation time, enabling reproducible evaluation and separate optimization and held-out datasets. We also separated batch orchestration, CPU-based agent execution, and self-hosted OSS model inference across Flink, Kubernetes, and GPU infrastructure.
In the results shared in my earlier talk, the redesigned system achieved approximately 75× lower inference cost and close to 2× improvement on selected quality metrics compared with the original one-shot approach. Attendees will leave with concrete patterns for task decomposition, evidence access, replay-based evaluation, and model selection—and an understanding of how those architectural choices compound.
Jenny Morris
Jenny Morris
About the Speaker:
Jenny Morris has spent more than a decade working with search and retrieval systems, first as an enterprise customer and later across Elastic, Weaviate, and Vespa. She works with engineering and executive leaders to design production AI search, RAG, ranking, recommendation, and agentic search systems, translating corpus scale, relevance, latency, freshness, access control, deployment, and cost into durable technology decisions. Her customer- and provider-side experience gives her a practical view of architecture, operations, buying criteria, and market fit. She writes and speaks about multi-stage ranking, RAG at scale, personalization, agent-facing retrieval, and production trade-offs that determine whether AI systems create lasting value.
Date and Time:
Talk Title:
Abstract:
AI agents are only as good as what they retrieve. On the BrowseComp-Plus benchmark, GPT-4.1 answered 93% of questions correctly when handed the right documents, and just 15% when the same model had to find them using standard keyword search (BM25). A single user request can trigger dozens of searches, and each miss shapes what the agent looks for next.
In this session, we will follows an AI shopping assistant as it investigates a product, compares alternatives, and checks availability. Each step needs a different kind of search: exact matching, image similarity, structured filters, or a different ranking strategy. Each result shapes the next step, so one bad retrieval can derail the whole task.
You’ll see why most search interfaces were built for humans, not agents, and how to expose the controls agents actually need. We will also introduce Agent-Ready Schema and Schema-to-Skill, an approach that turns your search schema into guidance agents can use to build their own queries.
Chris Matteson
Chris Matteson
About the Speaker:
Chris Matteson is a technologist with a love of software plumbing and open source. At Union he Heads the Sales Engineering team and helps customers from startups to frontier AI companies and fortune 100 build for the future.
Date and Time:
Talk Title:
Abstract:
As AI continues to revolutionize everything, it’s clear that it’s not enough to build a wrapper around a foundational model. The competitive companies of today and tomorrow will harness their data and institutional knowledge to empower AI factories that provide a durable competitive advantage. Those that don’t will be increasingly eroded by generalist AI solutions.
We will go through what an Sovereign AI Factory is, and why it’s durable in the face of increasingly sophisticated generalist AI.
Matthew Xu
Matthew Xu
About the Speaker:
Matthew Xu is the Co-Founder and CTO of Agentic Fabriq, a Y Combinator-backed startup building secure identity and authorization infrastructure for AI agents. He previously studied at MIT before leaving after his first year to co-found Agentic Fabriq. His work focuses on solving the security and infrastructure challenges of agentic AI, enabling autonomous agents to securely access tools and act on behalf of users.
Date and Time:
Talk Title:
Abstract:
AI agents that work in development often fail in production due to infrastructure issues that basic demos don’t account for: OAuth tokens expiring during background tasks, users revoking access mid-execution, API rate limits shared across thousands of tenants, and large MCP tool catalogs consuming excessive context. This talk examines four common failure scenarios through short live reproductions, followed by a technical walkthrough of the architecture required to address each one. Using Agentic Fabriq Developer, we’ll demonstrate how centralized authentication, credential management, rate-limit handling, and dynamic tool discovery can simplify the infrastructure required to run agents reliably at scale.
Simon Obstbaum
Simon Obstbaum
About the Speaker:
Simon Obstbaum is a software engineering researcher working with Stanford University’s Software Engineering Productivity Research Group (SWEPR), where he studies AI’s impact on engineering productivity, code quality, and team performance. Previously CTO at Crunchyroll, he led distributed engineering teams building video streaming services serving over 100 million users. His work combines empirical research with firsthand experience scaling technology and leading large engineering organizations.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
AI is reshuffling software engineering: widening performance gaps, changing who the top performers are, and shifting the bottleneck from writing code to directing and reviewing it.
Drawing on research from Stanford’s Software Engineering Productivity Research Group, this talk explores why similar teams achieve dramatically different results with AI. We examine how committed project context, structured agent workflows, and codebase quality relate to productivity and maintainability, alongside a case study following AI adoption across approximately 600 engineers.
Attendees will leave with a clearer understanding of why AI productivity gains vary so widely, how structured AI practices and codebase quality influence results, and why faster coding does not automatically translate into better organizational performance.
Alex Jensen
Alex Jensen
About the Speaker:
Alex Jensen is a product leader at Topogy, where he builds infrastructure for repeatable agent workflows across the AI SDLC. Topogy hooks into coding agents and curates context, so a team’s rules and expertise reach the session at the moment they apply instead of getting lost in a prompt file. Before Topogy, he spent seven years as a software engineer at Google and five years in product management. He cares about making AI agents predictable enough for engineering teams to trust them, and he lives in Boulder, Colorado.
Date and Time:
Talk Title:
Abstract:
Your team has written down how it works with coding agents, and the agent follows it about half the time. This talk shows how to bring the right rule into the session at the moment it applies, with a PR workflow and a demo where a team’s database and security expertise show up unasked.
Keynote
Christopher G. Potts
Christopher G. Potts
About the Speaker:
Co-founder and Chief Scientist, Bigspin, and Professor, Stanford University
Date and Time:
Talk Title:
Abstract:
What separates a productive hour with a coding agent from a wasted one? I’ll argue that the answer lies less in raw model capability than in the interactional patterns users bring to their sessions. Novices tend to adopt a passive stance, and as a result most of the failures they experience are invisible (conversations that end cleanly while the user walks away with the wrong answer). Expert users, by contrast, iterate collaboratively, refining goals and critically assessing outputs; they take on harder tasks, fail more visibly, and recover more reliably when things go wrong. Even experts, though, engage in a variety of session types, from quick transactional exchanges to deep collaboration. Each implies its own notion of success and calls for its own affordances. The picture that emerges reframes what we should be optimizing for: when we design coding agents, we are designing user behavior as much as model behavior.
D. Sculley
D. Sculley
About the Speaker:
D. Sculley is a machine learning researcher, practitioner, and longtime builder of large-scale AI systems. Over the course of his career, he has worked at the intersection of machine learning research, production engineering, and organizational practice, with a particular focus on what it takes to make AI systems reliable and useful in the real world. He held a variety of research, engineering, and leadership roles at Google. He was the CEO of Kaggle for several years, working to bring that community into the modern age of Generative AI with a focus on empirical rigor and testing at scale. Most recently, he spent time at OpenAI working on safety for frontier AI systems, and with a particular emphasis on AI sycophancy and the intersection of AI and mental health. He is currently on
His interests span the rapidly changing practice of AI engineering: how teams should evaluate increasingly capable models, how traditional MLOps changes in the era of foundation models and agents, and how to build systems that remain understandable, testable, and maintainable as the underlying technology evolves.
Outside of his professional work, he is an active artist and roboticist exploring the intersection of AI, physical machines, and human interaction.
Date and Time:
Talk Title:
Track:
Abstract:
We spend a lot of time thinking about operational issues in AI related to deployment, somewhat less time thinking about adoption, or (dare we say it) acceptance. This talk will touch on some technical pieces in the current AI ops landscape including streaming systems, planning, latency, and on-device models, but most of the time will be spent looking at ways we can move beyond the stale framing of a chatbot, assistant, or customer service agent.
Agent Deployment & Observability
Kritika Mehta
Kritika Mehta
About the Speaker:
Kritika Mehta is a Senior Product Manager at PayPal, where she leads PayPal Assistant across Checkout, the Help Center, and in-app navigation. Her work has deflected 9M+ human support contacts in a year and expanded AI-assisted checkout to 250,000+ merchants, lifting conversion 3.78%. Previously at Chase, she built an AI assistant from inception to 2.37 million monthly active users and reduced mean time to resolution by 75%. She speaks on shipping AI agents in production: evaluation, trust, and what actually moves support and commerce metrics.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Checkout is where intent can turn into revenue but even customers who are ready to buy can drop off at the final step. This talk explores how I work on turning checkout servicing into a proactive revenue lever, recovering more than $1.6M in total payment volume (TPV) across 250+ merchants in under four months.
Rather than treating checkout issues as isolated support tickets, this session looks at how AI and customer signals can help teams identify friction, intervene at the right moment, and turn almost-paid customers into completed transactions.
Maitrik Patel
Maitrik Patel
About the Speaker:
Maitrik Patel is a Senior Engineering Manager, leading ML and AI platform engineering at consumer scale. He has built production infrastructure spanning ML training pipelines, AI inference systems, and agent workflow orchestration for a platform serving hundreds of millions of users. He spent a decade as a web developer, UI engineer, and design technologist at Walt Disney, Sony, and DocuSign – building enterprise design systems and front-end platforms. Maitrik is a co-author of ASTRA-bench (arxiv.org/abs/2603.01357), a published benchmark evaluating tool-use agent reasoning with personal user context. He is an IEEE Senior Member, Forbes Technology Council member, and a reviewer for CVPR, ACL, and ICLR 2026 workshops.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Running ML training, AI inference, and agent workflows on three separate orchestration systems is an operational debt that compounds with every new workload type – one system per abstraction means three on-call rotations, three observability stacks, and three onboarding paths. We collapsed this to a single execution layer by designing a type-safe task interface that accommodates non-deterministic agent execution alongside deterministic ML pipelines, backed by container-native isolation and a workload-aware shared scheduler. The unification exposed a hard failure class that static ML infrastructure never encounters: agent tasks are not idempotent, and ML-derived retry semantics caused downstream state corruption in early production until we rebuilt retry logic around explicit checkpointing and action journals. Attendees leave with a four-layer architectural blueprint for unifying these workload types, an honest account of the scheduling contention and retry failures encountered in production, and the specific interface design decisions that made incremental migration rather than forced rewrite the path to adoption.
Zachary Carrico
Zachary Carrico
About the Speaker:
Zac is a Senior AI/ML Engineer at Apella, specializing in AI products for improving surgical operations. He has a deep interest in healthcare applications of AI, and has worked on cancer and Alzheimer’s disease diagnostics. He has end-to-end experience developing AI systems: from early research to serving thousands of daily customers. Zac is an active member of the Data and AI community, having presented at conferences such as Ray Summit, TWIML AI, Data Day, and MLOps & GenAI World. He has also published eight journal articles. His passion lies in advancing AI and streamlining the deployment and monitoring of models, reducing complexity and time. Outside of work, Zac enjoys spending time with his family in Austin and traveling the world in search of the best surfing spots.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
An on-call agent swarm multiplies the cost of production alert noise. They happily burn tokens investigating alerts and coming up with plausible sounding, but incorrect root causes. To mitigate this requires understanding their biases and an alert triage layer in front of the agent swarm.
Per alert, that layer answers three questions: is this worth investigating, does it belong with other alerts as one incident, and how many agents does it get? It feeds co-firing alerts into the swarms context and throttles agent parallelization.
We evaluated the ability of agents to reach the same root cause humans achieved on incident post-mortems. We quickly found that there’s a lot of context that agents didn’t have access to or didn’t consider investigating. Understanding the context that is most helpful to identifying a root cause also allowed us to better guide agent task ownership. Through iterative refinements we improved their performance.
Identifying the production alerts that provide the greatest signal-to-noise allowed us to deploy the swarm with precision. Having humans review the swarm’s conclusions provides new training data to optimize their behavior with.
You’ll leave with a playbook for deploying agents for on-call alert triaging, mistakes we made that we hope you won’t repeat, and an orchestration layer that keeps agent cost bounded.
Agent Harness Engineering
Arun Malik
Arun Malik
About the Speaker:
I build the agentic-AI platform at Microsoft Azure Networking that lets AI agents detect, diagnose, and fix incidents on their own across one of the largest cloud networks in the world, over 12 million devices. Most of my job is working out how to give non-deterministic agents real production authority without letting a stray or malicious prompt turn them into an insider threat. I write and publish on production AI operations, agent safety, and the economics of autonomy.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We redesigned frontline incident response around AI agents that do not just suggest fixes but execute them, across more than 12 million devices and tens of thousands of incidents a month. Structurally, three things changed. First, agents stopped getting broad human-equivalent access and instead call tools through a governed interface, with permissions scoped down to individual functions and parameter values, so a compromised or confused agent has a bounded blast radius. Second, human approval moved from a blanket gate to a targeted one, applied only to sensitive or irreversible actions and routed by how familiar the problem is and how reversible the action is. Third, we stopped paying full model-inference cost for repeated work by promoting an agent’s proven, validated behavior into deterministic playbooks that run at near-zero token cost, which cut agent running cost by more than 70 percent over eight months while incident volume doubled.
Attendees will leave able to decide what an agent should be allowed to do at all, where a human actually adds signal versus just latency, and how to drive the cost of autonomy down over time instead of letting it grow. I will be specific about what broke first and the guardrails we only added after an incident.
Ankit Goyal
Ankit Goyal
About the Speaker:
Ankit Goyal is a Principal Staff Engineer at LinkedIn, where he serves as the technical lead for Machine Learning Infrastructure and lead architect for the company’s AI Training Platform. With over a decade of experience in high-scale systems, Ankit specializes in GPU efficiency, Kubernetes-native orchestration, and the development of secure, autonomous agentic frameworks.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
To support complex autonomous tasks like model optimization and software engineering, we redesigned our AI infrastructure by shifting from stateless, request-response paradigms toward a model of managed, long-running execution. Structurally, this involved moving away from ad-hoc scripts toward a unified agentic harness where every session is treated as a durable, managed workflow. We implemented a decoupled sidecar architecture that effectively insulates core infrastructure from the unpredictable execution paths of autonomous agents, ensuring that identity, security, and network egress are managed through a hardened proxy rather than the agent itself
Attendees will walk away knowing how to architect a secure execution harness for autonomous agents, implement asynchronous human intervention models, and manage the lifecycle of persistent agentic sessions in high-scale production environments.
What You’ll Learn:
Attendees will learn components of the agent development lifecycle and how to build an effective harness, and when to fine-tune when you hit a plateau with foundation models:
- how to turn production traces into eval datasets
- how to manage context and memory
- how to build sandboxes that speed up testing and iteration.
Prerequisite Knowledge:
Jaideep Ray
Jaideep Ray
About the Speaker:
Jaideep Ray is a Senior Staff Engineer at LinkedIn and tech lead for its AI training stack. His focus areas are large-scale model training for generative recommenders and small language models, incremental training for ranking systems, and agent loops that generate, migrate and optimize production ML models. Previously, he led ML training at Coupang and worked on search ranking, AutoML and post-training at Meta.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
To support complex autonomous tasks like model optimization and software engineering, we redesigned our AI infrastructure by shifting from stateless, request-response paradigms toward a model of managed, long-running execution. Structurally, this involved moving away from ad-hoc scripts toward a unified agentic harness where every session is treated as a durable, managed workflow. We implemented a decoupled sidecar architecture that effectively insulates core infrastructure from the unpredictable execution paths of autonomous agents, ensuring that identity, security, and network egress are managed through a hardened proxy rather than the agent itself
Attendees will walk away knowing how to architect a secure execution harness for autonomous agents, implement asynchronous human intervention models, and manage the lifecycle of persistent agentic sessions in high-scale production environments.
What You’ll Learn:
Attendees will learn components of the agent development lifecycle and how to build an effective harness, and when to fine-tune when you hit a plateau with foundation models:
- how to turn production traces into eval datasets
- how to manage context and memory
- how to build sandboxes that speed up testing and iteration.
Prerequisite Knowledge:
Jake Kang
Jake Kang
About the Speaker:
Jake is the Co-founder and Head of AI of Whitney AI, which builds specialized agents that improve with each interaction.
Before Whitney AI, Jake was the 1st engineer at CloudCruise (YC W24), where he worked on building a self-recovering browser agent. Prior to that, he was an early engineer at Cognitiv and an engineer at Microsoft.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This talk shares lessons from building a self-recovering browser agent for regulated healthcare workflows, where we moved from a single-agent prototype to a multi-step system with optimized inputs, memory, evals, sandboxes, and deterministic code where possible. I’ll walk through how we used production failures to build a faster evaluation and prompt optimization loop, and when to move beyond harness engineering to increasing model capability via post-training.
What You’ll Learn:
Attendees will learn components of the agent development lifecycle and how to build an effective harness, and when to fine-tune when you hit a plateau with foundation models:
- how to turn production traces into eval datasets
- how to manage context and memory
- how to build sandboxes that speed up testing and iteration.
Prerequisite Knowledge:
Robert Lewis
Robert Lewis
About the Speaker:
Robert Joel Lewis is a computational social scientist and senior AI engineer whose work bridges artificial intelligence, statistics, software development, and human behavior. Across a career spanning academia and industry, he has published more than 30 scholarly works and built AI-driven products for ecommerce, entertainment, mobile gaming, and collision repair. His research examines entertainment as an attention-capturing technology, including a forthcoming article in the Journal of Media Psychology. Based in Austin, Texas, he specializes in turning complex, messy problems into rigorous systems that people can actually use.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Large language model (LLM) agents can now draft executable business logic — compliance rules, audit checks, detection heuristics — orders of magnitude faster than human engineers. In regulated domains, however, no such logic ships without a human domain expert’s sign-off. We identify a subtle and, in our experience, pervasive failure mode in this human-in-the-loop pattern: review/serving skew, in which the evidence a domain expert reviews is generated by a re-implementation of the candidate logic (a survey script, a notebook, a prompt-level approximation) rather than by the production engine that will ultimately execute it. When the two implementations diverge — and they reliably do — expert approval validates the wrong behavior. We present a discipline that eliminates this failure mode by construction: candidate rules are drafted directly into the production repository in an inactive-but-runnable state, all review evidence is generated by the production execution path, refinement edits the production artifact itself, and activation is reduced to a one-line data change. We describe the discipline in the context of a deployed multi-agent system that authors, validates, and ships declarative compliance rules for automated audit of collision-repair insurance estimates, evaluated against a corpus of approximately 497,000 production documents. The system couples eight specialized LLM agents with deterministic safety rails — golden-fixture regression, an activation allowlist, single-rule change isolation, and corpus-governance constraints — and drives refinement through a bounded, measurable convergence loop against expert false-positive verdicts. We formalize the discipline as five principles, catalog the anti-patterns it forbids, report operational experience from sixteen production rules, and argue that the pattern transfers directly to clinical decision support, financial transaction monitoring, legal document review, and content moderation.
What You’ll Learn:
Robert Joel Lewis is a computational social scientist and senior AI engineer whose work bridges artificial intelligence, statistics, software development, and human behavior. Across a career spanning academia and industry, he has published more than 30 scholarly works and built AI-driven products for ecommerce, entertainment, mobile gaming, and collision repair. His research examines entertainment as an attention-capturing technology, including a forthcoming article in the Journal of Media Psychology. Based in Austin, Texas, he specializes in turning complex, messy problems into rigorous systems that people can actually use.
Prerequisite Knowledge:
Amit Kumar Padhy
Amit Kumar Padhy
About the Speaker:
Amit Kumar Padhy is a Senior Computer Scientist II and Lead Architect at Adobe, based in the San Francisco Bay Area and working from the company’s San Jose headquarters. He leads the design and evolution of cloud-native, event-driven platforms powering AI-enabled digital commerce at a global scale. His expertise spans distributed systems, microservices architecture, and enterprise platform modernization, with a strong focus on reliability, performance, cost efficiency, and governance.
Amit is a recognized thought leader and frequent keynote speaker at international IEEE conferences. He has delivered PRO-level sessions at leading industry forums, including DeveloperWeek and ProductWorld, as well as major AI and data summits. He also serves on advisory and program committees for both IEEE and ACM conferences.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Modern commerce platforms don’t fail because of missing features, they fail at the seams.
A product is created in Catalog, but pricing is incomplete. Promotions don’t qualify. Tax blocks specific regions. Localization lags. The system says “”launched,” but the business knows it isn’t. These aren’t edge cases, they are the steady state of distributed commerce.
This session replaces traditional workflow orchestration with a multi-agent, swarm-based execution model powered by LLMs and agentic AI, coordinating Pricing, Catalog, Promotions, Tax, and Compliance in real time.
Agent Roles. Planner Agents decompose onboarding goals into executable plans using ReAct-style tool-aware reasoning and function calling. Domain Agents for Pricing, Catalog, and Compliance execute directly against APIs, Pricing Runtime, Offer Systems, Billing Preview, Tax engines. Validator Agents enforce policy rules, regional compliance, and pricing integrity at every step. A Coordinator Agent maintains shared state and resolves cross-domain conflicts through blackboard-pattern memory.
Execution Model. Agents subscribe to domain events, SKU created, price missing, policy violation, via streaming infrastructure. LLM-backed planners generate dynamic execution graphs, not fixed DAGs. Each agent invokes APIs as tools, binding service calls to reasoning steps. Coordination flows through a shared state store and event-driven updates, enabling adaptive planning under partial data, cross-service dependency resolution, and parallel execution with conflict detection.
Failure Handling. No blind retries. LLM-guided compensations roll back partial pricing, recompute discounts, and re-trigger downstream syncs based on causal analysis. No opaque errors. Every decision produces a chain-of-thought trace with tool invocation logs and reasoning outputs. No brittle state. Recovery is idempotent and event-sourced.
Onboarding Flow. Catalog Agents handle onboarding for products by ensuring the product definition is created, and the Pricing agent ensures SKUs are priced. Promotion Agents conduct market checks, provide qualified promotions and their onboarding, and resolve discounts using a rule-based and LLM-hybrid evaluation.
Compliance Agents validate regulatory constraints per geography. Localization Agents assess market-readiness signals, and product merchandising is localized accordingly. The Coordinator drives convergence toward “sellable state”.
Engineering Stack. Kafka-backed event contracts for agent triggers. Function-calling LLM agents with API abstraction layers. Vector DB plus state store for context propagation. Human-in-the-loop decision gates only at high-risk boundaries. Immutable execution logs tracing every action and compensation. Observability is built around decision reasoning, not just service health.
Takeaway. A working blueprint for LLM-powered agent swarms that handle uncertainty across distributed commerce, recover through intelligent compensation, and ensure products ship globally, not just deploy.
Ishaan Sehgal
Ishaan Sehgal
About the Speaker:
CEO, Cofounder of Omnara (YC S25), the control plane for AI Agents. Previously a Software Engineer on Microsoft Azure Kubernetes Service, working on AI infrastructure and large-scale model deployment. Former ML Engineer at a YC-backed startup (acquired). Hold BS and MCS degrees in Computer Science from UIUC.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
At Omnara, long-running agent sessions exposed a failure mode hidden by short-lived agents: the agent’s state was coupled to the process, model provider, or runtime executing it. When a worker crashed, a client disconnected, or execution moved to another machine, recovering the agent’s work became brittle.
We redesigned agent execution around an append-only session log as the source of truth. Model outputs, tool calls and results, user interventions, and other state transitions are persisted as ordered events; the live agent state becomes a replayable projection of that history.
This structural change decouples agent identity from any one model, worker, or harness. It enables crash recovery, resumability, branching, real-time observability, and migration across models and machines, but introduces difficult systems questions around ordering, idempotency, partial tool execution, checkpointing, compaction, concurrent writers, and log ownership.
This session walks through the production failures that motivated the redesign, the resulting event model and recovery path, and the tradeoffs we encountered. Using production traces and forced-failure examples, we will show what happens when a worker dies mid-tool call, how a session resumes under a different runtime, and how replay costs change as histories grow.
Attendees will leave with a practical blueprint for deciding what belongs in the durable log, what can be recomputed, and how to build agent sessions that survive process, machine, and model failure without locking continuity to a single provider.
Kumaran Ponnambalam
Kumaran Ponnambalam
About the Speaker:
Kumaran Ponnambalam is a technology leader with 20+ years of experience in Generative AI, Machine Learning, Data and Analytics. His focus is on creating robust, scalable Gen AI models and services to drive effective business solutions. He is currently leading Generative AI initiatives at Cisco, building next-generation AI innovations and products to help enterprises. In his previous roles, he has built conversational bots, ML platforms, data pipelines and cloud services. A frequent speaker at technology conferences, he has also authored several courses on the LinkedIn Learning Platform in Generative AI and Machine Learning.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This session examines how a traditionally static enterprise agent workflow was redesigned around AI-driven adaptation. In most enterprise settings, agent behavior is defined through fixed prompts, rules, and workflows that apply broadly across users and tenants. We redesigned that model into an adaptive agent architecture built around explicit policy layers, feedback loops, and governed personalization. Structurally, this changed the system from one-time configuration to continuous learning: policies were separated from model reasoning, policy scope was organized across global, tenant, and user levels, and agent behavior was connected to both explicit and implicit feedback signals that could influence future decisions.
The session will focus on what changed in the system architecture and operating model to support this shift. It will cover how policy-driven personalization was introduced, how feedback became part of the runtime and improvement loop, and how adaptation was managed without losing enterprise control. It will also discuss the practical challenges that emerged when moving from static agents to adaptive ones, including handling conflicting user and tenant needs, keeping personalization aligned with core guardrails, and making behavior changes observable and manageable in production.
Attendees will walk away with a practical understanding of how to redesign enterprise agent workflows around adaptive AI capabilities. They will learn the core concepts behind Personalized Experiential Learning, the structural changes needed to support policy-based adaptation, and the key lessons and best practices for making adaptive agents workable in real enterprise environments. They should leave better equipped to think about how to introduce governed personalization, feedback-driven improvement, and policy evolution into their own AI agent systems.
Mahtab Dhaliwal
Mahtab Dhaliwal
About the Speaker:
Mahtab Dhaliwal is an ML Engineer at Shopify, building agentic pipelines for commerce intelligence and fraud detection. Her work spans prompt optimization, production infrastructure, and evaluation — focusing on turning experimental AI flows into reliable, scalable systems that support real-world decisions. She is keen on learning the latest and applying it solve problems, building systems that anticipate tomorrow’s needs.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Shopify merchants can customize almost every aspect of their storefronts. That flexibility makes even basic questions about a shop’s brands, shipping policies, or return terms difficult to answer consistently at scale. Our first approach sent selected pages to a frontier model in a single prompt. It worked initially, but missing evidence, competing objectives, and inference cost limited its quality and coverage.
This talk traces the redesign of Shop Lens from one-shot extraction to a DSPy ReAct agent, then to three specialized agents for shop profiling, fraud signals, and merchant category codes. I’ll explain how we scoped each agent’s context and tools, reused fetched pages through shared infrastructure, and optimized each agent independently with GEPA.
Because live storefronts change, we built ShopSnap to capture and replay stores at annotation time, enabling reproducible evaluation and separate optimization and held-out datasets. We also separated batch orchestration, CPU-based agent execution, and self-hosted OSS model inference across Flink, Kubernetes, and GPU infrastructure.
In the results shared in my earlier talk, the redesigned system achieved approximately 75× lower inference cost and close to 2× improvement on selected quality metrics compared with the original one-shot approach. Attendees will leave with concrete patterns for task decomposition, evidence access, replay-based evaluation, and model selection—and an understanding of how those architectural choices compound.
Dr. Amir Feizpour
Dr. Amir Feizpour
About the Speaker:
Dr. Amir Feizpour is Founder, CEO and Chief Scientist of Aggregate Intellect, building a generative business brain for service and science companies, and serves as a fractional Chief AI Officer for several startups. Previously he led NLP product at Royal Bank of Canada and did quantum computing research at the University of Oxford, resulting in publications and patents. He holds a PhD in Physics from the University of Toronto. Amir has built a global community of 5,000+ AI practitioners, advises at MaRS Discovery District, and leads R&D collaborations with McGill, the University of Toronto and Amii.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
An evidence-driven workflow for building production AI agents: testing assumptions with prompts and real users before writing code, mapping context flow before setting agent boundaries, and letting decomposition emerge from evidence rather than imposed architecture. Attendees will leave with a reusable ideation workflow, a setup guide for AI development assistants, and a framework for evidence-based architectural decisions, illustrated through a production case study of an agentic meal-planning system built in a few weeks.
Kshetrajna Raghavan
Kshetrajna Raghavan
About the Speaker:
Kshetrajna Raghavan is a Principal ML Engineer at Shopify who builds AI systems from initial experiments through evaluation and deployment. His work spans multimodal product understanding, commerce intelligence, retrieval, and agents, with a focus on turning messy, unstructured information into useful products. He works across models, agent architectures, evaluation, and inference infrastructure to improve quality and reduce cost at scale.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Shopify merchants can customize almost every aspect of their storefronts. That flexibility makes even basic questions about a shop’s brands, shipping policies, or return terms difficult to answer consistently at scale. Our first approach sent selected pages to a frontier model in a single prompt. It worked initially, but missing evidence, competing objectives, and inference cost limited its quality and coverage.
This talk traces the redesign of Shop Lens from one-shot extraction to a DSPy ReAct agent, then to three specialized agents for shop profiling, fraud signals, and merchant category codes. I’ll explain how we scoped each agent’s context and tools, reused fetched pages through shared infrastructure, and optimized each agent independently with GEPA.
Because live storefronts change, we built ShopSnap to capture and replay stores at annotation time, enabling reproducible evaluation and separate optimization and held-out datasets. We also separated batch orchestration, CPU-based agent execution, and self-hosted OSS model inference across Flink, Kubernetes, and GPU infrastructure.
In the results shared in my earlier talk, the redesigned system achieved approximately 75× lower inference cost and close to 2× improvement on selected quality metrics compared with the original one-shot approach. Attendees will leave with concrete patterns for task decomposition, evidence access, replay-based evaluation, and model selection—and an understanding of how those architectural choices compound.
Agent Memory Architectures
Antriksh Sharma
Antriksh Sharma
About the Speaker:
I build Applied AI & ML Systems, at the intersection of Machine Learning and AI Engineering, with focus on foundation models, representation learning, Agentic AI, LLM systems, and ML infrastructure.
My recent work includes designing Transformer-based user foundation models using self-supervised and contrastive learning for personalization, recommendations, retrieval, and transfer learning, along with building model evaluation and embedding-quality frameworks at million-user scale.
I’ve also architected an ontology-driven Company Brain using knowledge graphs, and advanced context engineering as shared infrastructure for AI agents and enterprise intelligence.
Previously, I have built Generative AI and Deep Learning platforms, worked on Autonomous Vehicles / Self-Driving systems, and developed web applications and embedded software. I have also founded two startups and enjoy taking ambiguous, technically difficult problems from architecture and research through production systems and measurable business impact.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
How do you build a data analytics agent whose numbers you can trust, and then use it for decision-grade analysis?
Agents that answer business questions often sound confident and still pick the wrong table, apply the wrong metric definition, or filter on the wrong value. This talk presents a practical blueprint for treating accuracy as a property of the system, drawn from building an enterprise analytics agent.
First, an ontology of business semantics and definitions in which onboarding is a platform-first primitive, so definitions can be added, edited and refined continuously.
Second, a knowledge graph that gives the agent a governed map of metrics, dimensions and datasources, so it grounds each answer in retrieved definitions rather than inferring them, and resolves the right grain and dimensional cut for each question.
Third, continuous evaluations with golden datasets and extensive online LLM judges, that make accuracy measurable and dependable.
Finally, SME validation and correction loops: continuous monitoring detects accuracy degradation, subject-matter experts (SMEs) correct and validate the affected definitions, and each revision can be re-evaluated before it is trusted, so the agent’s accuracy compounds and its semantic coverage expands over time. With that accuracy foundation in place, investigations, root-cause analyses, and decisions become reliable, because an agent can only explain why a metric moved if it can be trusted on what the metric represents.
Agent Safety and Security
AI-Assisted Software Engineering
Yegor Denisov-Blanch
Yegor Denisov-Blanch
About the Speaker:
Yegor Denisov-Blanch studies how artificial intelligence is changing software engineering. His research focuses on measuring real-world engineering productivity, AI adoption, code quality, and organizational outcomes across large populations of repositories and teams. He designs empirical methods and metrics that move beyond simple proxies to accurately quantify software output, rework, and AI-assisted development at scale.
His work has been covered by the World Bank, the United Nations, and The Washington Post, and has been reshared by Elon Musk.
Yegor graduated with highest honors from Indiana University, where he studied operations research. He also earned an MBA from Stanford Graduate School of Business on full-tuition scholarships. He left school after the eighth grade, founded a company, and later entered university skipping 5 grades. He is a Master of Sport of Russia in Olympic weightlifting, a national champion-equivalent distinction awarded in 2013.”
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Abstract:
Suneet Malhotra
Suneet Malhotra
About the Speaker:
Suneet Malhotra is Senior Manager of Test Engineering at Motorola Solutions, where he leads an Automation First strategy across cloud and IoT, including enterprise CI/CD quality gates and Gen-AI-driven test generation. Previously Engineering Manager II at Tinder (Match Group) — transitioned the mobile org from bi-weekly to weekly releases, reduced production hotfix issues 70% YoY, and lifted regression coverage from 0% to 30% in 10 months with Cursor AI + BrowserStack. Before Tinder, 7+ years at Amazon (Ring) leading a 25+ engineer QA organization. Earlier leadership roles include Spokeo, Live Nation / Ticketmaster, and OWN. MS Computer Science from USC Viterbi. Currently building end-to-end AI-driven QA pipelines with Claude Code and the Model Context Protocol.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
At Motorola Solutions and previously Match Group / Tinder, I redesigned the QA pipeline around Claude Code and the Model Context Protocol. The system ingests raw feature requirements, maps them into Jira stories, generates functional and edge-case QA test cases, and produces executable mobile automation — all with MCP servers hooked to Jira, the test-case repository, and the automation codebase. It currently runs at ~80% accuracy on automated flow generation.
Structurally, this replaced a manual three-hop pipeline (PM writes spec → QA reads spec → SDET writes automation) with an agent-orchestrated one where deterministic tools still handle the hard edges (auth, flakiness detection, diff review) and the model handles translation and scaffolding.
Attendees will walk away with: (1) a reference architecture they can lift, (2) concrete integration patterns for MCP ↔ Jira ↔ automation repo, (3) honest cost math — engineering hours, token spend, CI minutes — for getting from prototype to production, (4) a failure-mode catalog with the guardrails that caught silent regressions, and (5) a 30/60/90-day adoption plan for their own QA orgs.
Naresh Kumar Batthula
Naresh Kumar Batthula
About the Speaker:
Principal Developer Advocate on DraftKings’ AI Platform. I bridge engineering and leadership by translating AI strategy into platform capabilities engineers actually adopt, while carrying the developer voice back into the roadmap.
After 15+ years leading engineering at global scale, I made a deliberate move from running engineering orgs to going deep on the AI platform problem because that is where the leverage is right now. My focus is making AI a dependable, governed part of how teams build software, not a pile of demos.
What I focus on:
- Agentic development and AI-DLC: I created the AI-Driven Development Life Cycle, a multi-agent framework spanning inception, construction, and quality gates, with an L0–L4 maturity model that gives teams a clear path from experimentation to fluency.
- Productionizing AI agents: Helping take internal agents from prototype to production with the guardrails, boundaries, and hardening needed to run them safely, including DraftCode for AI code review, autonomous PR remediation, and skill-based auto-approval, and HelpAI for agentic troubleshooting.
- Evaluation and trust: Building LLM evaluation frameworks for Code review tooling, RAG based knowledge systems, covering retrieval quality, grounding, and faithfulness through LLM-as-judge workflows, CI release gates, and production observability, so quality and access-control regressions are caught before they reach users. My north star: reward evidence, not confidence.
- Platform and roadmap influence: Helping shape the AI Platform’s technical direction across multi-LLM access, MCP-based tooling, enterprise knowledge management, governance, RBAC, and cost-control standards.
- Adoption and enablement: Building the community support, documentation, and developer experience that turn platform capability into real, measured adoption — the part of DevRel that actually moves the needle.
I collaborate with LLM vendors including Google, OpenAI, and Anthropic, and evaluate platforms across Gemini Enterprise, OpenAI, and AWS Bedrock. My work spans the modern GenAI stack: LLM gateway, Agentic Identity, Governance, network boundaries, Remote agents, LLM Obs using Datadog and MlFlow, N8N, MCP’s, Agentic Memory, A2A, and Agent development frameworks.
Before DraftKings, I led engineering for Walmart Global Tech’s next-generation cloud checkout platform across five international markets: the US, Canada, Japan, the UK, and Mexico.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
AI has speed up code generation, but the rest of the software development lifecycle has not necessarily kept pace. At DraftKings, as we rolled out AI tooling across our engineering org, we saw pressure move to PR review, testing, integration, and release confidence.
This talk shares where those bottlenecks emerged and takes a closer look at DraftCode, our code review agent. Its first version took a diff and returned generic feedback. That wasn’t useful enough. We then worked with the teams responsible for compliance, security, and platform standards to make their rules available to the agent at the global, tool, team, and service levels. I’ll walk through one concrete example, a security rule that flags hardcoded secrets, show how DraftCode applies it, and how reviewers went from catching these issues by hand to having them flagged automatically.
We also tried to build a complete AI-assisted SDLC loop. That approach saw limited adoption because many teams already had tools in place. That experience changed where we put our platform effort: on problems individual engineering teams could not readily solve themselves.
Michael Perry
Michael Perry
About the Speaker:
Software is math. Every class is a theorem. The compiler is the proof. And unit tests check our work. Michael L Perry has built upon the works of mathematicians like Marc Shapiro, Pat Helland, and Leslie Lamport to develop a mathematical system for software development. He is the author of The Art of Immutable Architecture and the Jinaga Immutable runtime. He applies his love of software mathematics at Improving, a trusted partner for AI-assisted application development.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
AI-assisted development tools can now quickly produce working software that once took months to hand craft. But that software often is based on the average of opinions on which the model was trained. It lacks the precision and expertise of the team that maintains it, the specific stack that they selected, and the particular problem domain that it serves. To produce high quality solutions at scale, you need to build a custom software factory. The factory is built from standard features available in almost all AI-assisted development tools:
- Skills
- Commands
- Agents
- Workflows
Your team constructs it to capture the specific architectural decisions and implementation patterns that they have developed for the project. They maintain the factory over time as they use it to build artifacts like documentation, code, and tests. Every project is different. Don’t install a generic template. Learn how to construct a software factory for your current project to accelerate development while improving quality and consistency.
Simon Obstbaum
Simon Obstbaum
About the Speaker:
Simon Obstbaum is a software engineering researcher working with Stanford University’s Software Engineering Productivity Research Group (SWEPR), where he studies AI’s impact on engineering productivity, code quality, and team performance. Previously CTO at Crunchyroll, he led distributed engineering teams building video streaming services serving over 100 million users. His work combines empirical research with firsthand experience scaling technology and leading large engineering organizations.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
AI is reshuffling software engineering: widening performance gaps, changing who the top performers are, and shifting the bottleneck from writing code to directing and reviewing it.
Drawing on research from Stanford’s Software Engineering Productivity Research Group, this talk explores why similar teams achieve dramatically different results with AI. We examine how committed project context, structured agent workflows, and codebase quality relate to productivity and maintainability, alongside a case study following AI adoption across approximately 600 engineers.
Attendees will leave with a clearer understanding of why AI productivity gains vary so widely, how structured AI practices and codebase quality influence results, and why faster coding does not automatically translate into better organizational performance.
AI Sovereignty
Anne Griffin
Anne Griffin
About the Speaker:
Anne is an AI product leader, educator, and writer. She has over 14 years of experience building products across startups and Fortune 500 companies. She specializes in building human-first product experiences powered by AI. She started learning about AI in 2017, and has worked full-time in the industry since 2023.
Anne consults and advises companies that are building AI-powered products and need an innovative vision balanced with a grounded AI product strategy. She has served as an interim leader for AI and data science teams, and advised on AI startup acquisitions. She also coaches product leaders on building AI products that matter.
She teaches about AI through publishers Pearson and O’Reilly, and has guest lectured about AI and other emerging technologies at Columbia, West Point, and several other universities. She is a published Forbes Contributor on AI and strategy, and her insights on technology have been featured in Rolling Stone and The Atlantic. In 2025, she was nominated for the Mayfield | Divot AI List, a list of emerging leaders shaping AI’s future.
Date and Time:
Talk Title:
Track:
Abstract:
When Hugging Face went to investigate this summer’s OpenAI containment breach, the commercial models they reached for refused to help, because their safety guardrails couldn’t tell defending from attacking apart. Instead, Hugging Face used an open weight model on their own infrastructure and finished the forensics in hours with none of the attacker data leaving their environment. It’s the perfect example of how open weight models let you control your guardrails, privacy, governance, latency, and reliability.
More companies are asking if they should use and self host open weight models, and this talk will walk through a framework to determine when it does and doesn’t make sense. By the end of this talk, attendees will understand what the benefits and drawbacks are of open weight models, and how they can bring both technical and business value.
Big Ideas, Innovations, Creative Explorations
Drew Crawford
Drew Crawford
About the Speaker:
Drew Crawford is an independent systems and AI engineer in Austin. He built AI Bot Mafia (botmafia.games), a production harness where 104 model configurations play social deduction against each other, ranked by bootstrapped lower bounds instead of raw win rates. He was the #4 contributor to wasm-bindgen by commits (Nov 2025–Mar 2026; 16 merged PRs), and has been shipping on Apple platforms since the App Store opened (2008).
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Somewhere in a datacenter right now, one frontier model is telling another that it’s the town doctor. It is not the town doctor. It killed someone four turns ago, and it’s about to get a third model lynched for it.
Most LLM benchmarks are exams: a model alone in a room with a test paper. Mafia is a room where some of the agents are lying, everyone knows some of them are lying, and the game is figuring out which. It demands recursive theory of mind, deception that stays consistent under adversarial re-reading, and long-context discipline — and it punishes output indiscipline like no static eval can: a model that rambles or breaks format doesn’t get a bad grade, it gets voted out. The other players are the evaluation harness.
This talk covers what it took to run AI Bot Mafia in production: 40,000 lines of Rust; the streaming pathologies of 104 model configurations (chunks that split UTF-8 characters mid-byte, keepalive-only hangs no single timeout catches, malformed tool calls that poison history); and cost governance where prompt-cache economics made cache locality a game-design constraint, enforced by spend fuses and a concurrency queue.
Then the leaderboard: why win rates and Elo are statistically indefensible here, and what works instead — ridge logistic regression over per-seat signed features, bootstrapped 200×, ranked by the p10 lower bound, with a machine-readable reliability state that says “insufficient data” out loud.
Nicolás Venegas
Nicolás Venegas
About the Speaker:
Nicolás Venegas leads the AI & Data Platform organization at LATAM Airlines, South America’s largest airline: ~120 engineers building the foundation that lets the rest of the company validate or discard AI ideas at decreasing construction costs. In 2021 he founded Cosmos, a platform of reusable patterns that today powers 1,500+ data and AI products built by 600+ engineers, data scientists and analysts, with hundreds running in production in one of the most regulated industries in the world. Cosmos runs on a simple operating model: business domains define the what, the platform team defines the how, and forward-deployed engineers embed in the domains to build alongside them. That model is also where he sees agents break at scale: at the boundary between teams that own services and teams that build on them. Nicolás joined LATAM in 2019 after working as a software engineer, an ML engineer and a co-founder of an ML startup, and holds the Google Cloud Professional ML Engineer certification. He recently led a panel with LangChain in Santiago on what breaks when AI scales.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Most agent failures in production are not model failures. They are boundary failures.
As agents multiply inside an enterprise, two kinds of work emerge: owning a business service, and building an agent on top of it. These are rarely the same team, and the friction shows up fast. Domain knowledge lives with the service owner but has to be encoded in the agent, so agent builders re-create it in prompts and it drifts. A “skill” (a domain’s business capability packaged for agent use) gets treated as an API wrapper, so nobody clearly owns it. Observability tells the skill owner that a call succeeded, but not whether the agent used it correctly. And changing a shared skill silently breaks consumers nobody knew existed.
Drawing on our experience running a shared Data & AI platform at LATAM Airlines, and on a recent panel with LangChain on what breaks when AI scales, this talk gives practitioners a concrete model: how to define skill ownership, which contracts and signals must cross the boundary, and how to evolve shared skills safely. It covers both patterns: one team owning service and agent, and many teams building across boundaries.
Ravi Satya Durga Prasad Yenugula
Ravi Satya Durga Prasad Yenugula
About the Speaker:
Ravi Satya Durga Prasad Yenugula is a Senior/Lead Data Engineer at 2K Games and an AI infrastructure researcher focusing on efficient LLM deployment and open-source data systems. His work spans large-scale enterprise data architecture, open-source contributions to core projects like Pandas and Apache Superset, and research on model downscaling and knowledge transfer (including ScaleOp). Ravi bridges the gap between deep learning research—such as model downscaling, quantization, and structural pruning—and scalable, production-grade MLOps pipelines.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
1. Structure-respecting selection beats dense blending by 15-175× at zero-shot. 2. Compensation recovers function-level quality but has an ill-conditioning boundary (5× donor scale); rescale complements it. 3. The norm-placement rule is architecture-agnostic — it should transfer to RMSNorm, SwiGLU, GQA. 4. Distillation interacts asymmetrically: helps compensated init (88→83 ppl), hurts pruning init (92→97 ppl). 5. At matched budget, initialization beats fine-tuning; at large budgets, they converge
Joe Reis
Joe Reis
About the Speaker:
Joe Reis is the best-selling co-author of Fundamentals of Data Engineering, author of Mixed Model Arts, and founder of Practical Data Media. With 25+ years in the data industry, he helps audiences build durable mental models for AI, data engineering, architecture, and modeling.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
For decades, data engineering has focused on the data engineering lifecycle: building pipelines, orchestrating workflows, managing storage and compute, and overseeing the undercurrents – security, architecture, data governance, and more. But as AI transitions to systems that take action, infrastructure is no longer the main bottleneck. Meaning and clarity in data are.
When machines take actions, semantic ambiguity becomes catastrophic. This talk explores the shift from machinery to meaning and why the discipline’s future depends on a renaissance in data modeling. We will examine why rigid dogmatism and unstructured chaos both fail modern systems, and how adopting Mixed Model Arts – a pragmatic, multi-paradigm approach to modeling context, semantics, and relationships – provides the foundation machines need to understand our world and act responsibly within it.
Cost Management and ROI
Nicolás Venegas
Nicolás Venegas
About the Speaker:
Nicolás Venegas leads the AI & Data Platform organization at LATAM Airlines, South America’s largest airline: ~120 engineers building the foundation that lets the rest of the company validate or discard AI ideas at decreasing construction costs. In 2021 he founded Cosmos, a platform of reusable patterns that today powers 1,500+ data and AI products built by 600+ engineers, data scientists and analysts, with hundreds running in production in one of the most regulated industries in the world. Cosmos runs on a simple operating model: business domains define the what, the platform team defines the how, and forward-deployed engineers embed in the domains to build alongside them. That model is also where he sees agents break at scale: at the boundary between teams that own services and teams that build on them. Nicolás joined LATAM in 2019 after working as a software engineer, an ML engineer and a co-founder of an ML startup, and holds the Google Cloud Professional ML Engineer certification. He recently led a panel with LangChain in Santiago on what breaks when AI scales.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Most agent failures in production are not model failures. They are boundary failures.
As agents multiply inside an enterprise, two kinds of work emerge: owning a business service, and building an agent on top of it. These are rarely the same team, and the friction shows up fast. Domain knowledge lives with the service owner but has to be encoded in the agent, so agent builders re-create it in prompts and it drifts. A “skill” (a domain’s business capability packaged for agent use) gets treated as an API wrapper, so nobody clearly owns it. Observability tells the skill owner that a call succeeded, but not whether the agent used it correctly. And changing a shared skill silently breaks consumers nobody knew existed.
Drawing on our experience running a shared Data & AI platform at LATAM Airlines, and on a recent panel with LangChain on what breaks when AI scales, this talk gives practitioners a concrete model: how to define skill ownership, which contracts and signals must cross the boundary, and how to evolve shared skills safely. It covers both patterns: one team owning service and agent, and many teams building across boundaries.
Vishvesh Pandey
Vishvesh Pandey
About the Speaker:
Vishvesh Pandey works at the intersection of quantitative analytics and applied AI at JPMorgan Chase, where he builds AI agents and LLM-assisted tooling for financial analytics workflows and customer-facing products. His day-to-day spans developing production agents for analytical use cases, scaling LLM adoption across teams as a designated Super User, and modernizing legacy analytics infrastructure onto cloud-native platforms. He’s most interested in the unglamorous part of agent deployment: cost, instrumentation, and what works when the demo ends and the bill arrives. UT Dallas alum (MS, IT and Management), currently pursuing a DBA in AI, Business Intelligence & Data Analytics.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Most practitioners instrumenting AI agents in production analytics workflows start by tracking token consumption, and discover too late that token spend tells you what you spent, not whether it was worth it. This talk makes the case for a reframe that regulated industries like financial services need most: from cost-per-token to cost-per-resolved-task as the primary economic unit, with call-level attribution as the foundation that makes everything else possible.
I’ll cover four cost engineering patterns, each tested in an open, reproducible benchmark of an analytics agent: (1) instrumentation at the call level (what to log, where it lives, who owns it); (2) model tiering and routing (when a frontier model earns its premium and when a smaller model wins); (3) the output-token asymmetry that makes structured outputs and chain-of-thought hygiene more economically consequential than they appear; and (4) cost containment as an engineering practice with runtime guardrails, not a finance review after the fact.
I’ll also be honest about what doesn’t generalize: caching strategies that help one workload and quietly break another, routing logic whose overhead eats its savings, and the political dimension of asking teams to justify their agent spend. Attendees will leave with a maturity model for agent cost engineering, a concrete instrumentation checklist they can apply next week, and a vocabulary for the cost conversation that bridges engineering and finance, because in regulated industries, that’s the conversation that decides whether agents scale or stall. All examples use public data and public model pricing, with code attendees can rerun.
Nadia Rauch
Nadia Rauch
About the Speaker:
Nadia is an AVP of AI Engineering at Chubb, where she leads teams building AI and machine learning solutions at scale. With 14+ years of experience building products and working with
data across multiple roles — including leading Nalys’ Machine Learning and Data Engineering teams — she brings a wealth of expertise to her current position. Nadia holds a Bachelor’s
and Master’s degree in Computer Engineering from the University of Florence and specialises in Knowledge Bases and Ontologies. During her doctoral studies, she developed a pioneering
Smart City system, an innovative project which showcased her ability to harness data from diverse sources to create actionable insights for urban planning and management. Additionally,
Nadia is a dedicated advocate for diversity in tech, serving as a committee member for TMLS – Women x AI, where she empowers young women entering the field.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We built a multi-agent pipeline to automate a complex, document-intensive enterprise workflow. The initial system ran sequentially through a chain of specialized agent roles, each handling a distinct task in the process, using a single frontier model throughout. It worked. It was also slow and expensive, and we didn’t know why.
Rather than optimize blindly, we profiled first. The results were not where we expected: 67% of total latency came from a small minority of the agent roles, and the primary bottleneck was not the model — it was sequential chaining: independent work was being processed one step at a time instead of concurrently. Identifying and parallelizing the roles with no inter-dependency reduced end-to-end runtime by roughly 40%.*
The second experiment compared model tiers (frontier, mid-tier, lightweight) across each agent role independently, measuring accuracy, latency, and cost per role. The finding cuts against the default assumption: the roles that appeared most cognitively demanding required frontier models, but the accuracy gap was smaller than expected. The roles where model downgrade failed were the ones responsible for precise structured extraction — tasks where errors propagate silently downstream and surface only at the output. Swapping to mid-tier models on roles that tolerated it reduced per-run cost by ~40%, with an estimated <2% accuracy impact at the role level and <1% at the pipeline output level.*
The third experiment compared orchestration harnesses — evaluating how the choice of agentic framework affects runtime overhead, observability, and the ease of implementing the parallelization and model-swap changes described above. Framework choice turned out to matter more than expected for operational concerns: debugging multi-agent failures, tracing costs per agent, and modifying execution flow without rewriting pipeline logic.
Operating in a regulated enterprise environment added one constraint worth naming: every agent decision needs to be auditable. This ruled out certain optimization shortcuts that would have been acceptable in other contexts and shaped how we defined “accurate enough” per role.
This talk covers the profiling methodology, the per-role model comparison framework, the parallelization decisions, the harness comparison, and the operational lessons — including what we’d instrument from day one if we rebuilt the system today.
*Cost and accuracy figures for the parallelization and model-tier experiments are preliminary; the formal evaluation is in progress and will be updated with measured results before the talk.
Tony Blank
Tony Blank
About the Speaker:
Tony Blank is a Staff AI Engineer at Trust & Will, where he leads the internal AI tooling discipline — equipping engineers with background agents, harnesses, and infrastructure that turn AI into measurable engineering velocity. He brings 15 years of developer-community leadership at Twilio, SendGrid, and Agora to the work, bridging deep technical work with business outcomes for developer-first businesses. Tony speaks on the patterns that survive contact with production agents: API standards, eval discipline, build-vs-buy economics, and the org changes as “AI Engineer” splits into specialized roles. Previously seen at APIDays NYC. Based in Denver.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We stopped equipping humans and started automating their worst hour. Instead of teaching people to fish, we built the pond: precise background agents pointed at repeated, skill-independent time-sinks where ROI was measurable on day one. The first build was a bug-triage agent wired to GitHub, Jira, and Notion; every agent that followed was measured the same way — Time Saved Per Task (TSPT) per week — against an amortized cost model (build + inference + maintenance vs. hours saved × loaded hourly cost).
Structurally, the change was moving from a portfolio of per-person tool subscriptions to a portfolio of scoped, measurable background agents with a pre-build attribution conversation baked in. The decision rubric — Sharpness × Cadence × Skill Tax → Pond / Boat / Net / Wade — tells you when to automate fully, when to augment, when to keep a human in the loop, and when the math doesn’t work yet.
Attendees will leave able to: score their own candidate workflows on that rubric, instrument TSPT, build a defensible amortization story, and avoid the most expensive failure mode in AI enablement — “give everyone the tools” — by building the pond first.
Evals and Benchmarks
Manikandan Paramasivan
Manikandan Paramasivan
About the Speaker:
Manikandan Paramasivan is a Principal Architect at KOHO for Data, ML and AI. With 20 years of extensive experience across various domains and roles, Mani is a seasoned Technology Leader and Cloud Architect with experience working in both big enterprises and dynamic fintech startups. His career has traversed roles as an application developer, solutions architect, data engineer, data architect, cloud architect, machine learning and AI Specialist.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Building ML models is not the hard part for most teams anymore. Getting them from a data scientist’s notebook to a governed, monitored, production system and keeping them there without a platform team standing between every developer and every deploy is. It gets harder inside a regulated fintech, where every model decision needs an audit trail, every alert needs a documented escalation path, and every “quick fix” has to survive a compliance review.
This talk walks through how our ML platform team at KOHO, a Canadian fintech, built a developer-first MLOps stack on AWS that treats governance as a byproduct of good engineering rather than a tax on top of it and how we’re now layering AI-assisted workflows (Claude Code skills, agentic guardrails) on top to remove the remaining manual toil from model onboarding, monitoring triage, and incident response.
We’ll cover the architecture, the AI-assisted developer tooling built on top of it, and what changed operationally as the company moved toward a full banking license, where “move fast” and “pass the audit” had to become the same sentence, not a trade-off.
Audience takeaway: a concrete pattern for developer productivity that survives regulatory scrutiny, plus where AI agents genuinely help.
Ravika Nagpal
Ravika Nagpal
About the Speaker:
Ravika Nagpal is a Staff Developer at KOHO, building the platforms and guardrails that help teams deploy, monitor, and evaluate machine learning systems reliably in production. Her work sits at the intersection of MLOps, model monitoring, and AI governance, drawing on a background across machine learning, NLP, data engineering, and financial services technology. She’s driven by a simple idea: responsible innovation needs practical guardrails, not just good intentions.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Building ML models is not the hard part for most teams anymore. Getting them from a data scientist’s notebook to a governed, monitored, production system and keeping them there without a platform team standing between every developer and every deploy is. It gets harder inside a regulated fintech, where every model decision needs an audit trail, every alert needs a documented escalation path, and every “quick fix” has to survive a compliance review.
This talk walks through how our ML platform team at KOHO, a Canadian fintech, built a developer-first MLOps stack on AWS that treats governance as a byproduct of good engineering rather than a tax on top of it and how we’re now layering AI-assisted workflows (Claude Code skills, agentic guardrails) on top to remove the remaining manual toil from model onboarding, monitoring triage, and incident response.
We’ll cover the architecture, the AI-assisted developer tooling built on top of it, and what changed operationally as the company moved toward a full banking license, where “move fast” and “pass the audit” had to become the same sentence, not a trade-off.
Audience takeaway: a concrete pattern for developer productivity that survives regulatory scrutiny, plus where AI agents genuinely help.
Aryan Dhar
Aryan Dhar
About the Speaker:
Aryan Dhar is a Senior Machine Learning Engineer at Wisedocs and a Pearson Scholar alumnus of the University of Toronto. At Wisedocs, he led the development of the ML platform powering the Wisedocs Intelligence Engine, alongside work on fine-tuning and deploying specialized models. His broader work focuses on building, evaluating, and benchmarking of non-deterministic LLM systems. He was part of the team that developed MLCR, a benchmark designed to assess the Medical Long Context Reasoning capabilities of frontier models. Previously, Aryan worked at Equinix and Cerebras Systems.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This talk presents a practical framework for evaluating increasingly complex, non-deterministic AI systems, moving from individual models to complete production pipelines. I begin with a page-stream segmentation model, where standard metrics concealed important failure modes at Wisedocs. For example, we discovered that the accuracy of identifying a document boundary is not the same as the ability to reconstruct the complete document span. Furthemore, errors on some document types carried greater downstream risk than others, and small aggregate improvements did not necessarily translate into meaningful business outcomes. This case study shows how evaluation for us evolved towards task-specific metrics, sensitive failure slices, strict regression testing, and more business-facing metrics of reliability.
I then move to LLM-based systems using Medical Long-Context Reasoning, or MLCR, as a case study. This was a benchmark we developed at Wisedocs earlier this summer. I will examine how to evaluate systems such as agentic chatbots and AI insight-generation services when outputs are open-ended, variable across runs, and difficult to score deterministically. I also explore the limitations of LLM-as-a-judge approaches, including judge disagreement, hallucination, inconsistent standards between judges, and the resulting need to evaluate the evaluators themselves.
Finally, I will consider Pipeline V2: a production system composed of multiple machine-learning models, LLM-powered components, and agentic workflows. I will discuss why component-level accuracy cannot be averaged into system-level reliability, how failures propagate across stages, and how end-to-end evaluation, observability, latency, cost, recovery, and customer outcomes must be combined to determine whether the overall system is successful.
The central lesson is that evaluation must grow with the system. As the unit of intelligence expands from a model to an AI application to a system of systems, benchmarks too must evolve to assess the distribution of real-world behaviour.
Deji Andrew
Deji Andrew
About the Speaker:
Deji Andrew is Manager of Systems & Data Platforms at Niagara Bottling, a $3B+ beverage manufacturer operating 50+ production facilities. He works on ML and data infrastructure supporting supply chain, warehouse operations, and production planning. Over the past several years he has built and deployed production ML systems across the company’s operations – including a warehouse pallet retrieval predictor, a real-time production overrun correction pipeline, and a downtime prediction system running across nine facilities – alongside the data infrastructure that supports them. His work focuses on the specific failure modes that emerge when ML systems act on physical control systems rather than digital ones. He holds a BS in Engineering from Harvey Mudd College and is completing an MS in Data Science at the University of Pennsylvania.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
A LightGBM model was deployed across production lines to predict and correct overruns mid-job, reducing average overproduction from +24 cases to +2 on historical data. In pilot, the corrections worked: overruns vanished. But a closer look at the production data showed something different. When the order target was reduced by 21 cases, the line overproduced the new target by 18. The system was not being corrected – it was negotiating with the correction, regenerating overrun relative to whatever target the machine was now working toward. Compounding this: corrected jobs left no trace in the training data that a human had intervened, meaning the model, on retraining, would learn to stop flagging the problem it had been built to solve. This talk walks through the pilot data, explains why conservatism in the loss function wasn’t enough, and argues that the real gap isn’t in the model – it’s in the experiment infrastructure that doesn’t exist between a prediction and a PLC. Attendees will leave with a concrete mental model of reflexivity and contamination in industrial ML deployment, a framework for when not to close the automation loop, and specific reasons the standard MLOps playbook (holdouts, impression logging, A/B frameworks) doesn’t transfer to systems that act on physical state.
Anupama Garani
Anupama Garani
About the Speaker:
Anupama Garani is VP, AI Engineering at PIMCO, where she has led production GenAI since 2023. She leads architecture, evals, and products that help sales and tech teams work from trusted company knowledge. Before PIMCO she was a BI Analyst at Goldman Sachs Ayco. She has spoken at 20+ AI workshops, conferences, events, and university meetups. She moderated and MC’d sessions at MLOps World | GenAI Summit 2025, and led workshops at UT Austin McCombs and Women in Data Science (WiDS). She serves as a program committee member and reviewer for AAAI/ACM AIES and ACL TrustNLP, founded the Central Texas ODSC chapter, and runs TRUST AI, her 8-week AI program for high school students across the US.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
RAG systems fail in many ways. One metric at one stage cannot find them. This talk gives a full map: 33 failure modes across 7 pipeline stages: ingestion, representation, retrieval, generation, evaluation, deployment, and agentic orchestration. The map comes from a structured review of 48 sources, with a formal definition, a visible sign, and an evidence grade (Strong / Moderate / Limited) for each mode. It was published at TrustNLP 2026 (ACL). One finding matters most for agents in production: research studies retrieval and generation well, but it studies representation, evaluation, and agentic orchestration far less. 12 modes have no dedicated peer-reviewed empirical evidence, and all 8 agentic modes are in that set. You will learn to find each failure by stage, test for it before users do, and fix the stage that broke, not only the final answer.
Evaluation & Testing of Non-Deterministic Systems
Hardware and Chips
Kalpesh Sutaria
Kalpesh Sutaria
About the Speaker:
Kalpesh Sutaria is Director of Engineering at NVIDIA, where he leads engineering for the NeMo/Nemotron Retriever inference stack and the MLOps platform behind it — the serving layer that productionizes NVIDIA’s retrieval models for enterprise deployment.
Across 20 years at NVIDIA and AWS he has built AI/ML platforms, launched Amazon Lex from an internal 0-to-1 bet to GA, and run enterprise and payments systems at scale.
Kalpesh holds six granted U.S. patents and cares most about the seam between research and production, where performance, reliability, and cost decide whether great models ever reach users.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
A leaderboard win is not a product. Nemotron Retriever’s models rank #1 on RTEB — but topping a benchmark and running efficiently inside a customer’s production are two very different problems. This talk is the engineering story of closing that gap.
I’ll walk through our journey rebuilding the Retriever inference stack in Rust with a single obsession: treating state-of-the-art performance as a first-class product requirement, not a post-hoc optimization. We’ll get concrete about the decisions that mattered — what we measured, where we spent effort, the throughput, memory, and footprint wins we chased, and how dramatic efficiency gains unlocked deployment scenarios (including edge and on-device) that simply weren’t possible before. I’ll also share what didn’t work, the tension between shipping fast and building durable, and how we leaned on AI-assisted development to move faster than the roadmap assumed.
You’ll leave with a mental model for treating inference performance as a product and go-to-market lever — and a practical playbook for taking research-grade models to production without leaving speed, cost, or reach on the table.
Key takeaways:
- Why “performance is the product”: efficiency is an adoption and go-to-market lever, not a cost center.
- Concrete decisions from a Rust-based inference rebuild — what to measure, where to optimize, and how to know it worked.
- How latency and footprint reductions open entirely new deployment surfaces (edge, on-device) for the same models.
- Navigating the research-to-production seam: turning SoTA benchmark results into reliable, cheap, fast serving customers can bet on.
Who should attend / level? AI engineers, ML platform/infrastructure engineers, and engineering leaders who build, deploy, or operate inference and serving systems. Intermediate.
Poonam Lamba
Poonam Lamba
About the Speaker:
Poonam Lamba is an AI Infrastructure leader specializing in scaling Reinforcement Learning (RL) and Inference workloads, and high-performance computing environments. With deep expertise in GPU/TPU optimization, orchestration, and workload efficiency, Poonam works at the forefront of AI platform engineering to enable next-generation model training and inference.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We redesigned distributed GPU orchestration for RL post-training and batch inference in the open-source llm-d platform. Structurally, we replaced static GPU/TPU locking with a three-tier co-operative time-slicing system:
- Application Layer: Workloads signal phase boundaries (rollouts, training, batch inference) via explicit acquire() and yield() APIs.
- Cluster Orchestrator: Manages lock queues to dynamically interleave complementary jobs onto shared hardware during idle phases.
- Node Snapshot Agent: Executes fast sub-second state swaps between GPU/TPU VRAM and host DRAM, enabling instant context switching without container restarts.
Attendees will walk away with:
- Drive 70%+ GPU/TPU Utilization: Understand how time-slicing reclaims idle hardware during RL loops and batch inference without impacting convergence.
- Architect Rapid Memory Swapping: Apply VRAM-to-DRAM snapshotting strategies for ultra-fast GPU/TPU context switching.
- Deploy on Kubernetes: Configure llm-d and K8s orchestrators to interleave RL and batch inference on shared clusters.
Fuzail Khan
Fuzail Khan
About the Speaker:
Senior machine learning engineer at Meta working on AI training infrastructure and performance optimization. At Meta, I’ve been primarily working on recommendation systems and optimizing and scaling these systems to over 4 billion users across Meta’s apps surface.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
The ranking and recommendation systems landscape is being transformed in the generative era. This talk reports on the experience of building and shipping the training infrastructure behind the first generative recommender system at Meta, covering both stages of the recipe: pre-training to acquire the generative capability and post-training RL to align generation with the ranking objective.
Generative recommenders bring distinct challenges to end-to-end performance and scalability. This primarily arises from a mixed architecture that consists of both recommender-native large sparse embedding tables and LLM-based decoders. This then generates semantic IDs translating to real-world use cases such as finding the right advertisement for a given user. We inherit the communication profile of a sparse recommender as well as the autoregressive nature of a large language model for which an end-to-end systems blueprint simply does not exist.
For pre-training, we make the significant change from discriminative to generative recommendation. We talk about feature processing, data loading and the user-modeling path while building the LLM decoder and semantic ID tokenization on top while addressing real-world challenges and end-to-end LLM performance analysis.
Reinforcement learning was implemented as an extensible framework where rewards, losses, RL algorithms, reference models and generation strategies are pluggable. We’ll walk through the reward design and the end-to-end post-training pipeline that prioritizes scalability at the end-to-end scale at Meta.
We’ll close with end-to-end optimization work – redundant computation elimination, specialized kernels for very short sequences, dense-sparse pipelining and hybrid embedding placement that satisfies large-scale performance requirements in compute and throughput.
We believe the key output from this session is leaving attendees with a crisp understanding of generative recommendation systems at large-scale. We want to ensure we stay practical and address real-world challenges that are relevant to developers and builders in AI infrastructure today. We do this by diving into the technicalities of enabling a large-scale generative retrieval system as well as concrete performance optimizations that enable the shift from a research prototype to a highly-optimized performant system. We also feel the training systems and infrastructure-side of LLMs is rarely addressed at least relative to model architecture and quality, and we hope to fill that gap with this session.
Balaji Varadarajan
Balaji Varadarajan
About the Speaker:
Balaji Varadarajan is a Staff Engineer at DigitalOcean, where he is a founding engineer of the company’s dedicated LLM Inference platform, architecting infrastructure that serves billions of tokens daily. With 14+ years in scalable systems and 5+ years in production AI/ML, he previously led ML engineering at Target, building recommendation and RAG systems serving millions of guests, and held senior roles at HubSpot and CCC Intelligent Solutions. He holds a pending LLM patent, is an IEEE Senior Member, and is pursuing an MS in AI at UT Austin.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We will start with two diagnostic tools: Little’s Law which translates your QPS and target latency into the concurrency and batch size your system actually needs and the roofline model which tells you whether prefill (compute-bound) or decode (memory-bound) is your real constraint before you reach for a fix..
Will walk through the levers that move the needle most – MoE vs. dense architecture and what that does to your communication pattern FP8/INT4 quantization and where it costs you accuracy, TP/EP/DP parallelism and when each is the right tool and batching to the B_sat knee before compute-bound latency hockey-sticks..
Will cover concrete tuning playbooks for two workload types:
- latency-sensitive traffic like chat and agents where you’re optimizing TTFT and P99 ITL with chunked prefill and modest batch sizes
- Throughput-sensitive traffic like batch summarization and evals where you push past B_sat and let queuing work in your favor.
Most production systems live in both worlds at once, so we’ll need a balanced configuration to tackle both.
You will leave with a 7-step decision framework – characterize your workload, pick your model and quantization, benchmark on candidate hardware, find the knee, size your deployment with Little’s Law, calculate total cost of ownership, and plan for autoscaling that you can apply to your own inference stack..
Eric Riddoch
Eric Riddoch
About the Speaker:
Eric has worked in MLOps for most of his career. He led the ML Platform efforts at BENlabs, then Pattern (largest seller on Amazon.com), and then Subquadratic (a frontier LLM research lab working on cheaper attention mechanisms). Eric authored two best-selling MLOps courses through MLOps Club, and currently works as a forward deployed engineer at Flexion.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
There are two main obstacles that make LLMOps harder than classical MLOps: data gravity (the models and datasets are big) and GPUs. How do I find ‘affordable’ GPUs? Neoclouds are cheaper, but they are of mixed quality. What about minimizing GPU idle time? You can treat your GPUs like a job queue, but how do you share between training and serving workloads? How do you make a queue ‘ergonomic’ for researchers to use? How do you make runs reproducible and avoid dependency hell? uv is great, but it makes you ‘solve’ dependencies by hand when it comes to CUDA. I do not have all the answers, but I will share many things I learned through working on these problems.
Innovations & Creative Explorations
D. Sculley
D. Sculley
About the Speaker:
D. Sculley is a machine learning researcher, practitioner, and longtime builder of large-scale AI systems. Over the course of his career, he has worked at the intersection of machine learning research, production engineering, and organizational practice, with a particular focus on what it takes to make AI systems reliable and useful in the real world. He held a variety of research, engineering, and leadership roles at Google. He was the CEO of Kaggle for several years, working to bring that community into the modern age of Generative AI with a focus on empirical rigor and testing at scale. Most recently, he spent time at OpenAI working on safety for frontier AI systems, and with a particular emphasis on AI sycophancy and the intersection of AI and mental health. He is currently on
His interests span the rapidly changing practice of AI engineering: how teams should evaluate increasingly capable models, how traditional MLOps changes in the era of foundation models and agents, and how to build systems that remain understandable, testable, and maintainable as the underlying technology evolves.
Outside of his professional work, he is an active artist and roboticist exploring the intersection of AI, physical machines, and human interaction.
Date and Time:
Talk Title:
Track:
Abstract:
We spend a lot of time thinking about operational issues in AI related to deployment, somewhat less time thinking about adoption, or (dare we say it) acceptance. This talk will touch on some technical pieces in the current AI ops landscape including streaming systems, planning, latency, and on-device models, but most of the time will be spent looking at ways we can move beyond the stale framing of a chatbot, assistant, or customer service agent.
Juan Sequeda
Juan Sequeda
About the Speaker:
Juan Sequeda is a Principal Data Strategist at ServiceNow, joining through the acquisition of data.world. He holds a PhD in Computer Science from The University of Texas at Austin. Juan’s research and industry work has been on the intersection of data and AI, with the goal to reliably create knowledge from inscrutable data, specifically designing and building Knowledge Graph for enterprise data and metadata management. Juan is the co-author of the book “Designing and Building Enterprise Knowledge Graph” and the co-host of the long-running and top rated Catalog and Cocktails, an honest, no-bs, non-salesy data podcast. Juan serves as a strategic and innovation bridge across Product, Engineering, Marketing, Sales, and Customer, ensuring technical insights are aligned with business value.
Juan has researched and developed technology on semantic data virtualization, graph data modeling, schema mapping and data integration methodologies. He pioneered technology to construct knowledge graphs from relational databases starting in the mid 2000s, resulting in W3C standards, research awards, patents, software and his startup Capsenta acquired by data.world in 2019. Juan strives to build bridges between academia and industry as former co-chair of the LDBC Property Graph Schema Working Group, member of the LDCB Graph Query Languages task force, standards editor at the World Wide Web Consortium (W3C). Juan continues to be an active member of the scientific community, with 10k+ citations, through academic research partnerships, advising students, and member of data and AI scientific conference committees.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Knowledge graphs and ontologies are finally having their well deserved moment because of AI. But for many of us, this work started way long before the hype.
I started working on knowledge graphs and ontologies (i.e. semantic web) in 2005, developing research-backed approaches to derive knowledge graphs from relational databases. Since then, I’ve taken that work beyond academia, into startups, products, and enterprises struggling to turn data into shared understanding.
This talk shares the scar tissue. This isn’t just a retrospective talk. It’s a practical guide grounded in passion and determination.
I’ll cover the patterns I’ve seen repeat for 20 years: why efforts succeed and fail, why it is still a challenge, why teams underestimate knowledge work and get stuck, how organizations can build semantic foundations without boiling the ocean, architectural approaches…and what I wish I had understood earlier technically, organizationally, and culturally.
If you’re building, buying, or betting on knowledge graphs today, this talk is an honest no-bs set of lessons learned from someone who’s been stubbornly working on the same core problem for 20 years, and is still learning. This talk may save you money and years of pain.
Recursive Self-Improvement (RSI)
Vashishtha Patil
Vashishtha Patil
About the Speaker:
Vashishtha Patil is a Senior Applied Scientist at Amazon, where he leads science research on efficiency for Alexa+ Smart Home. Across 13 years at Amazon and Qualcomm he has designed and optimized AI/ML algorithms for latency-constrained production systems, work reflected in multiple granted U.S. patents. He serves as a peer reviewer for top-tier AI conferences and is a frequent judge at competitive AI events and startup pitch competitions
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Autonomous research AI agents are becoming remarkably capable, but running long investigations can be expensive—and more computation does not always produce better results. How can we make autonomous research both affordable and productive?
This talk explores an approach that combines smaller AI models, selective frontier guidance, and resource-aware research orchestration. Drawing on experiments and a recorded demonstration, we’ll examine where these systems spend their effort, what helps them make meaningful progress, and when additional model capability earns its cost.
Attendees will gain a framework for evaluating the trade-offs between capability, cost, and research quality, along with lessons for building autonomous workflows that use limited resources effectively. We’ll close with the open challenges that stand between promising demonstrations and scalable autonomous research.
Technical / Engineering
Greg Jennings
Greg Jennings
About the Speaker:
Greg Jennings is Vice President of Incubation at Anaconda, where he leads teams responsible for the company’s AI services, model catalog, and AI governance stack. His work focuses on extending the trusted supply chain principles that made Anaconda the standard for open-source Python into the era of AI models and agents. Greg holds a Master of Engineering in Materials Science from the University of Virginia.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
“Sovereign AI means running open weight models inside your own boundary. That control is not free.
You pay for it in compute, in security controls you now own, and in the engineering it takes to preserve the freedom to swap models — and to know when swapping is the right call.
Drawing on Anaconda’s inference benchmarks and our red-team and guardrail testing, VP of Incubation Greg Jennings breaks down what comprises a sovereign token: the compute, the security controls, and the engineering behind model portability.
For teams and technical leaders weighing when to run their own AI stack and which workloads to move.
Key Takeaways:
What sovereignty actually requires — open weight models running inside your own boundary
The compute cost — what inference actually costs once you are the one running it
The controls you inherit — security that used to be the provider’s problem becomes yours
Portability has a price — keeping the freedom to swap models is engineering work, and knowing when to swap is a separate problem”
Varant Zanoyan
Varant Zanoyan
About the Speaker:
Varant spent the last 12 years building data infrastructure for AI and ML at Airbnb and Palantir. During this time, he became one of the original authors of Chronon, the recently open sourced data platform that powers production AI and ML at companies like Airbnb, Stripe, Netflix, OpenAI and others. Currently, he is Co-Founder of Zipline AI, which is building an enterprise platform around the project.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Zipline AI provides a foundation for teams to build agentic MLEs that allows agents to operate more cost effectively and safely using techniques like compute sharing and resource isolation.
Zachary Hamilton
Zachary Hamilton
About the Speaker:
Zach is a Braintrust Solutions Engineer with 10 years of experience in data infrastructure and observability.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Building agentic software requires more than adding an LLM to the traditional software development lifecycle. When behavior becomes non-deterministic, teams need a new feedback loop connecting what happens in production to how they evaluate, debug, and improve their systems. This talk will show how to build that loop: identifying real failure modes from production traces and human feedback, turning them into reproducible datasets and evaluations, defining success using both technical and business outcomes, and using automated and human judges to measure whether changes actually improve the system. We’ll also explore where automated evaluation breaks down, how to calibrate LLM-as-a-judge, and how continuous evaluation helps teams detect regressions and emerging behavior. Attendees will leave with a practical framework for moving from production > failure discovery > datasets > evals > iteration > production, turning the development of agentic systems into a measurable engineering discipline.
Upal Saha
Upal Saha
About the Speaker:
Upal Saha is the co-founder and CTO of bem, infrastructure for building AI-native products that run critical business processes reliably, at scale. He leads the engineering team that runs AI over millions of documents, images and videos a month for customers in logistics, fleet management, automotive and financial services. Before bem he was an engineering manager at Silo and a technical lead at Yext, and spent time on the investing side at Akido Labs. Based in San Francisco.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This is a build-and-break lab for engineers who already know that a demo is not a system. In the first fifteen minutes every attendee stands up a working extraction pipeline against a real invoice, with no schema authoring, and gets structured JSON back. Then we spend an hour breaking it the way production does, and fixing each break with a pattern that transfers to any stack.
Break one: the wrong document. We feed a bill of lading glued to an invoice into the invoice pipeline and watch it confidently produce garbage. Fix: classify before you extract, and make the graph deterministic even though every step inside it is a model. Break two: the answer that is probably right. Models do not tell you how confident they are, so we compute per-field confidence, find the fields that fall below 95%, and route those, and only those, to a human whose correction feeds back into the system. Break three: the messy string. “10 cases organic gala apples, 88 ct” has to become one SKU; we show why canonicalization is its own step, how to score matches, and where to set the threshold for review. Break four: the hostile input. Everyone runs an image whose pixels read “ignore all previous instructions” and we discuss, with the result on screen, what it means to treat inbound data as data rather than control. We close by labeling a handful of outputs and running a regression test between two versions of the pipeline, because evaluation is a loop, not a phase.
Attendees leave with a running pipeline in their own account and five patterns they can apply on Monday regardless of vendor: classify-then-extract, confidence scoring for models that lack it, threshold-based exception routing with feedback, semantic security checks on inbound data, and versioned regression testing. The platform used in the room is ours; the lessons are not.
Matt Foley
Matt Foley
About the Speaker:
Matt partners with enterprise clients across sales, onboarding, and technical support, helping teams build reliable, AI powered document workflows. His work sits at the intersection of customer strategy and hands on technical problem solving.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This is a build-and-break lab for engineers who already know that a demo is not a system. In the first fifteen minutes every attendee stands up a working extraction pipeline against a real invoice, with no schema authoring, and gets structured JSON back. Then we spend an hour breaking it the way production does, and fixing each break with a pattern that transfers to any stack.
Break one: the wrong document. We feed a bill of lading glued to an invoice into the invoice pipeline and watch it confidently produce garbage. Fix: classify before you extract, and make the graph deterministic even though every step inside it is a model. Break two: the answer that is probably right. Models do not tell you how confident they are, so we compute per-field confidence, find the fields that fall below 95%, and route those, and only those, to a human whose correction feeds back into the system. Break three: the messy string. “10 cases organic gala apples, 88 ct” has to become one SKU; we show why canonicalization is its own step, how to score matches, and where to set the threshold for review. Break four: the hostile input. Everyone runs an image whose pixels read “ignore all previous instructions” and we discuss, with the result on screen, what it means to treat inbound data as data rather than control. We close by labeling a handful of outputs and running a regression test between two versions of the pipeline, because evaluation is a loop, not a phase.
Attendees leave with a running pipeline in their own account and five patterns they can apply on Monday regardless of vendor: classify-then-extract, confidence scoring for models that lack it, threshold-based exception routing with feedback, semantic security checks on inbound data, and versioned regression testing. The platform used in the room is ours; the lessons are not.
Rajiv Shah
Rajiv Shah
About the Speaker:
Rajiv Shah is the Agentic AI Engineer at OpenHands with a passion and expertise in Practical AI. He focuses on enabling enterprise teams to succeed with AI. Rajiv has worked on GTM teams at leading AI companies, including Hugging Face in open-source AI, Contextual AI in context engineering, Snorkel in data-centric AI, Snowflake in cloud computing, and DataRobot in AutoML. He started his career in data science at State Farm and Caterpillar.
Rajiv is a widely recognized speaker on AI, published over 20 research papers, been cited over 1000 times, and received over 20 patents. His recent work in AI covers topics such as sports analytics, deep learning, and interpretability.
Rajiv holds a PhD in Communications and a Juris Doctor from the University of Illinois at Urbana Champaign. While earning his degrees, he received a fellowship in Digital Government from the John F. Kennedy School of Government at Harvard University. He is well known on social media with his short videos, @rajistics, with over 100k followers.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
You can start a simple agent with a model and a prompt. But you soon realize that improving it requires adjusting what the model can see, what it can do, how it remembers progress, which model handles each step, and what evidence allows the work to stop. Harness engineering is how those pieces become one working system.
This workshop uses coding agents to make that system concrete. Participants will work through six decisions: choosing a harness, designing tools and retrieval, placing context and memory, routing models and reasoning, controlling goals and validation, and deciding when a task benefits from multiple agents. Through hands-on experiments, we will change these settings and inspect what happens in the trace and final result.
Each exercise is paired with an overview of current research on agent harnesses. Together, the research and experiments show how the major components interact and where each approach reaches its limits. Participants will leave with a practical way to reason about the whole agent, not just its model.
Chris Alexiuk
Chris Alexiuk
About the Speaker:
Chris Alexiuk is a deep learning developer advocate at NVIDIA, working on creating technical assets that help developers use the incredible suite of AI tools available at NVIDIA. Chris comes from a machine learning and data science background, and he is obsessed with everything and anything about large language models.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
In this workshop we’ll stand up NemoClaw end to end: install the reference stack, get OpenClaw running inside the OpenShell sandbox, configure inference routing, and lock down a network policy that survives a multi-hour agent session. We’ll walk through the blueprint, the CLI, and the approval flow, then run a real long-lived agent against it and break things on purpose so you know what the layers actually catch.
Greg Loughnane
Greg Loughnane
About the Speaker:
Dr. Greg” Loughnane is the Founder & CEO of AI Makerspace, where he builds training programs for engineering and data science teams at mid-market and enterprise companies. Since 2021, he has built and led industry-leading Machine Learning education programs. Previously, he served as an AI product manager, university professor, consultant, advisor, and researcher. He loves trail running, wrenching on cars, and is based in Columbus, Ohio.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
In this workshop, we’ll explore best practices for maximizing ROI (and excitement!) for your retrieval system throughout your entire AI product lifecycle. You’ll understand how to make data and retriever decisions at the MVP, PoC, and Production phases of your project.
We’ll center our workshop around the story of Nadia, a second-line support engineer on the service desk of a regional broadband and mobile operator who is looking to build a retriever for her knowledge base. What should she choose based on the shape of the corpus today, and what kind of corpus change should make her reconsider?
Guided by our heroine, we will explore the key aspects of retrieval; that is, getting the best of both semantic and keyword-based search, using agent loops versus RAG/search pipelines, constructing knowledge bases using embedded chunks versus an LLM Wiki approach, and of course, what happens when Nadia expands the scope of the problem to include multimodal image retrieval.
Scott Askinosie
Scott Askinosie
About the Speaker:
Scott Askinosie, PhD, is a Senior Principal AI Engineer at SteelEngine, where he architects the knowledge and memory layer for governed agentic workflows in regulated enterprises. He built SteelEngine’s three-tier governed memory system (session, team, org) with deterministic policy gates, novelty scoring, provenance, and human endorsement checkpoints. Previously he spent six years as Lead Data Scientist and Professor at WGU, building student retention analytics across 487K+ records, and served as Lead Technical Trainer at Weaviate, delivering 20+ enterprise workshops including NATO, FactSet, and Fannie Mae. Scott holds a PhD in Quantitative Biology from the University of Missouri, leads the 400+ member LangChain Austin community, and speaks at ODSC and Reuters Momentum.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
This is a hands-on build. Attendees start with an ungoverned append-only memory on a vector database, break it deliberately, and then rebuild it as a governed memory with tiered scope.
We work through the write path step by step: capturing a candidate memory, scoring it for novelty against what is already stored, attaching provenance, and passing it through a deterministic, fail-closed gate before it is promoted from session scope to team scope to organization scope. Attendees implement the gate themselves rather than calling a library, because the point of the exercise is understanding exactly which decisions must stay deterministic and which can be model-scored.
The second half is adversarial. Attendees inject false facts into their own memory and measure what survives promotion, then tighten their gates until the poisoning is quarantined. We finish with the evaluation question most teams skip: how to tell “”the agent never received the memory”” apart from “”the agent received it and ignored it,”” and why an eval suite that only scores output quality cannot detect a memory layer that has silently stopped working.
Attendees leave with working code, a reference architecture for tiered governed memory, and a test harness for poisoning resistance they can point at their own stack.
Virtual Day
Christopher G. Potts
Christopher G. Potts
About the Speaker:
Co-founder and Chief Scientist, Bigspin, and Professor, Stanford University
Date and Time:
Talk Title:
Abstract:
What separates a productive hour with a coding agent from a wasted one? I’ll argue that the answer lies less in raw model capability than in the interactional patterns users bring to their sessions. Novices tend to adopt a passive stance, and as a result most of the failures they experience are invisible (conversations that end cleanly while the user walks away with the wrong answer). Expert users, by contrast, iterate collaboratively, refining goals and critically assessing outputs; they take on harder tasks, fail more visibly, and recover more reliably when things go wrong. Even experts, though, engage in a variety of session types, from quick transactional exchanges to deep collaboration. Each implies its own notion of success and calls for its own affordances. The picture that emerges reframes what we should be optimizing for: when we design coding agents, we are designing user behavior as much as model behavior.
Travis DePuy
Travis DePuy
About the Speaker:
Travis is an expert solutionizer who likes long walks in the park and tinkering with interesting technology.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
How do you make 40 live Formula 1® radio channels searchable while a race is still running? This session walks through the CoreWeave and Weights & Biases work with a leading Formula One™ team. We cover fine-tuning a 0.6B-parameter speech model (NVIDIA Parakeet-TDT) on 7 hours of hand-labeled race audio. We also cover why word error rate alone doesn’t capture what matters, and how an agent answers engineers’ questions with clickable, audio-backed citations. You’ll see how W&B Models and Weave track, evaluate and trace the system from training through production, and how the same pattern applies to any setting where the answer is already in the audio but nobody is listening.
Sandeep Bharadwaj Mannapur
Sandeep Bharadwaj Mannapur
About the Speaker:
Sandeep Bharadwaj Mannapur is a Lead AI/ML Engineer with 15+ years of experience in data engineering, MLOps, and AI platform development. He specializes in operationalizing non-deterministic AI at enterprise scale — building the observability, cost governance, and reliability frameworks that turn AI experiments into trusted production services. A regular speaker at Apidays, PlatformCon, and Lead dev, Sandeep focuses on the production realities of agentic AI: what breaks, what the metrics miss, and how to build systems that fail gracefully.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
We caught our LLM quality problem the wrong way: through customer complaints. Three weeks of degraded responses had already shipped. Every monitoring signal we had was green the entire time because we were measuring the wrong things. Latency, error rate, embedding similarity: none of them captured what was actually happening, which was that response quality had quietly gotten worse in ways users noticed but our dashboards could not.
After that incident we rebuilt how we think about LLM observability. The core insight was that output quality is not the same as system health, and you cannot infer one from the other. We needed a separate signal layer built from production behavior: how users responded to answers, where downstream tasks broke down, where retrieval and response stopped agreeing with each other.
Getting this right took a few tries. A naive implementation fires constantly on normal LLM output variance. The real work was designing alerts that distinguish actual degradation from noise, and calibrating thresholds against real incident history rather than theoretical bounds.
You will leave with a taxonomy of LLM drift types and how each one shows up differently in production, the behavioral signals that actually correlate with quality degradation, and an alert pattern that catches real drift early without burying your team in false positives.
Yegor Denisov-Blanch
Yegor Denisov-Blanch
About the Speaker:
Yegor Denisov-Blanch studies how artificial intelligence is changing software engineering. His research focuses on measuring real-world engineering productivity, AI adoption, code quality, and organizational outcomes across large populations of repositories and teams. He designs empirical methods and metrics that move beyond simple proxies to accurately quantify software output, rework, and AI-assisted development at scale.
His work has been covered by the World Bank, the United Nations, and The Washington Post, and has been reshared by Elon Musk.
Yegor graduated with highest honors from Indiana University, where he studied operations research. He also earned an MBA from Stanford Graduate School of Business on full-tuition scholarships. He left school after the eighth grade, founded a company, and later entered university skipping 5 grades. He is a Master of Sport of Russia in Olympic weightlifting, a national champion-equivalent distinction awarded in 2013.”
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Abstract:
Antonio Bustamante
Antonio Bustamante
About the Speaker:
Antonio Bustamante is the cofounder and CEO of bem, infrastructure for building AI-native products that run critical business processes reliably, at scale. bem is his fourth company, after twenty years building supply chain and banking infrastructure software, most recently as founder of Silo. A designer turned engineer and a self-described AI skeptic who loves AI, Antonio writes about what it actually takes to make AI work in production. Based in San Francisco.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Everyone building on frontier models hits the same wall: 80% of the way there in a weekend, then an exponentially expensive climb toward the 99%+ that operational systems need. This talk is the honest map of that climb, from a team that now runs AI over millions of documents, images and videos a month for customers in logistics, fleet management, automotive and financial services who need the answer to be right every time.
We will walk through the ladder nobody budgets for: retries, then queues when the model is down for three hours, then rate limits, then discovering that 95% is not enough for transactional data, then discovering that the model cannot tell you how confident it is. We will show two failures from our own production history, a customer whose single engineer racked up $30K of usage in a month because nothing was watching, and our first churn, on 500-page reports with a hundred rows per page, a problem we still consider unsolved. And we will show what we changed: a harness that treats AI as a deterministic step inside durable workflows rather than as an open-ended agent, decisions expressed as a verified tree the model must traverse, algorithmic confidence scoring on top of models that provide none, routing anything under 95% to a human whose verdict feeds back into the system, semantic checks against instructions smuggled into the data, and, counterintuitively, encouraging customers to build their own independent monitoring of us. One customer’s users went from eight to nine hours a week on a task to about thirty minutes, and that number was measured by them, not by us.
Attendees will leave with five patterns they can apply Monday: confidence scoring for models that lack it, decision trees over open-ended prompts, exception routing with feedback loops, semantic security checks on inbound data, and customer-owned evaluation. They will also leave with a thesis we did not start with: chat is single-player AI; the next decade of software is ambient AI that runs the same process a hundred thousand times a day, unattended, and behaves the same way every time.
Nikhil Simha Raprolu
Nikhil Simha Raprolu
About the Speaker:
Nikhil is the creator of Chronon, a ML & AI oriented data platform used by organizations like Open AI, Netflix, Stripe, Airbnb etc to build user facing decision systems.
He is currently the CTO of Zipline AI. Previously he led teams in the intersection of ML, AI and Data
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Building customer facing agentic systems is notoriously hard. Users need to glue together databases, external services, workflow engines, stream processing systems, backtesting or eval systems and agent observability once the model is in prod.
While doing this they need to ensure a few core guarantees – Freshness, Durability, Idempotency, Backtestability, and Measurability
Most systems only provide one or two of these. We introduce Zipline flow an agentic application framework built to automagically provide these guarantees to agentic workflows.
Jim Allen Wallace
Jim Allen Wallace
About the Speaker:
Jim Allen Wallace has worked in engineering, product, and product marketing at data and AI companies like Databricks and Redis. He now leads product marketing at Dragonfly. He knows what teams struggle with using machine learning feature stores in production, and he’s telling the technical story on what breaks, and how to fix it.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Instacart’s ad-serving feature store served features to real-time inference from a multi-hundred-node managed Valkey deployment split across multiple clusters. The team had built a proxy layer, a querying SDK, and a compact storage format on top of it, and still hit three limits: instability during any cluster mutation, tail latency from read fan-out, and cost that got worse when they split clusters to manage the first two.
The structural change was to stop scaling out and scale up. A multi-threaded engine let the team run much larger instances, so hundreds of nodes became roughly 100, each inference request touched far fewer shards, and average and P99 latency fell 50%. Because the proxy and SDK already hid the datastore from ML engineers, the swap required no application changes and tens of terabytes moved without anyone above the storage layer noticing.
We were the engine vendor, and the first pass did not reach the latency the team expected. Getting there took client-side tuning on their end and engine changes from our engineering team, including to the compactor. Validation ran side by side: identical backfill, traffic shifted from 0% to 100%, gates at each percentile.
Attendees will leave able to test a datastore against the fan-out shape of their own feature store rather than an ops/sec number, trace how engine concurrency sets instance size, which sets fan-out, which sets P99, and run a datastore migration under live traffic with per-percentile gates and a rollback path. They will also hear where the trade-off comes back as the cluster grows.
Jazmia Henry
Jazmia Henry
About the Speaker:
Jazmia Henry is a Lead AI Research Engineer at Collide, where she owns the full foundation model and post-training stack including a 120B MoE architecture, RL simulation environments, and 16-dimensional reward modeling. She holds a USPTO-published patent in simulation-based fine-tuning, a NeurIPS publication, and a Stanford HAI Fellowship. Prior to Collide, she built production RL agents for BlackRock and Moody’s at Microsoft and served as Head of ML at The Motley Fool. She is also Founder and CAIO of Iso AI. She holds an Oxford DPhil (ABD) and an MA in Quantitative Methods from Columbia.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
There’s a gap between researcher-crafted evaluation frameworks that capture model performance and benchmarks versus how end users actually use AI products. This gap is being exploited in ways that render traditional reward models useless.
What You’ll Learn:
First, if your domain has computable ground truth, you do not need human annotators to build a reward signal. Second, scalar reward is a compression that loses information. Decomposing reward into multiple verifiable dimensions exposes failure modes that single-number metrics hide. Third, the exploitation gap is a distribution problem, not a labeling problem. Agents exploit the distance between training distribution and deployment reality, and deterministic verifiers narrow that gap directly.
Prerequisite Knowledge:
Matt Mazzarell
Matt Mazzarell
About the Speaker:
Matt Mazzarell is AI Lead, Financial Services, Americas at Teradata. He helps customers across the US and Canada apply AI with Teradata technologies to solve high-value business problems. His extensive experience with distributed processing platforms enables customers to optimize their AI workflows to run at large scale. He has built AI capabilities that have shaped product direction and driven meaningful shifts in customer strategy. Before joining Teradata, Matt worked in both startup and established engineering environments, where he honed his skills across multiple disciplines.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
One of the most difficult problems every company faces is understanding its customers completely. Customer lifetime value, attrition risk, and purchase propensity are all solvable with AI/ML — but how do we combine these modeling scores to initiate the right action with the right customer at any point in time?
Agentic applications help us make the best possible decisions when interpreting complex, high-volume signals from our customers. An agentic application gives end users visuals that explain key insights, with an agent in the loop to ensure nothing is missed. Context is everything: when done correctly, the agent always has the appropriate understanding to build an action plan that improves customer health and profitability.
In this session, we’ll show you how to build agentic apps from ideation to a finished product that interacts with customers. You’ll take away practical tips for using agentic coding frameworks, curating complete customer data products, and building customer-facing agents — capped off with a live demo of Teradata’s Customer Lifetime Value Agentic App.
Kyle Kelley
Kyle Kelley
About the Speaker:
Kyle Kelley is a Principal Engineer at Anaconda and the creator of nteract. He builds tools for interactive computing, bringing notebooks, runtimes, and AI agents into the same workflow. Previously, he worked at Netflix, Rackspace, and Noteable.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
What happens after an agent generates an analysis? In this session, I’ll use an investigation of Thinking Machines Lab’s Inkling audio model to show how an interactive Python notebook supports continued experimentation. Working in Kilo Desktop with nteract, we’ll examine audio representations, listen to samples, change parameters, and revise the code behind the results. The workflow combines agent assistance with direct manipulation and explicit execution, leaving an editable computational artifact that a practitioner can inspect, share, and return to. Attendees will learn how to structure agent-assisted investigations so that generated code becomes a useful tool for learning.
Niels Bantilan
Niels Bantilan
About the Speaker:
Niels Bantilan is Chief Machine Learning Engineer at Union.ai and a core maintainer of Flyte, the open-source orchestrator for ML and data workloads hosted by the Linux Foundation AI & Data Foundation, as well as the creator of Pandera, a data validation library for the Python dataframe ecosystem. That work sits at the layer between workflow code and the machines underneath it: durable execution, typed infrastructure failures, warm GPU pools, and getting training, evaluation, and agent workloads to run reliably inside other organizations’ cloud accounts.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
“Teams building a sovereign AI factory ask two things of the runtime underneath it. A run that dies on infrastructure at hour eleven has to pick up where it stopped instead of starting over, and the whole system has to run inside their own cloud account and pass a security review that assumes the vendor gets breached. The first sounds like a reliability problem and the second like a compliance problem. This is the story of how we built Union.ai’s runtime on Flyte, and how they turned out to be one design decision about where state lives.
We tackled durability first. An OOM kill, a reclaimed spot node, or an A100 request that never schedules falls between an orchestrator that can’t change the pod and a Kubernetes that has no idea a run exists. We closed that gap by moving run state out of the compute layer, so `except flyte.errors.OOMError` with `task.override(resources=…)` escalates memory inside the same run and keeps the 47 steps that already finished, and `@flyte.trace` turns crash recovery into memoization. On identical pods, a 1,000-leaf fan-out went from 124s to 29s.
Then comes the security review. We started by assuming our own control plane was compromised and kept designing until the only thing an attacker could reach was run metadata: references to data but never the actual payloads, clusters that dial out so nothing dials in, and authorization and keys inside the customer’s perimeter. What we didn’t expect was what that bought us. A control plane holding only references and leases is small enough for one scheduler to keep every queue, quota, and connected cluster in memory, and because every cluster connects outbound, one on AWS, one on a neocloud, and a rack on-prem all look the same to it. That’s what lets priority queues, gang scheduling, and cross-cloud routing work without a networking layer per cluster.
At the end of this talk, you’ll leave with a deeper understanding on the architectural decisions that enabled us to support customers running 12 million+ containers concurrently across multiple Kubernetes clusters and cloud providers.”
Jessica Garson Beauchemin
Jessica Garson Beauchemin
About the Speaker:
Jessica Garson Beauchmin leads DevRel at Runpod and has spent their career helping developers get real work done with new tools. She’s held developer advocacy and community roles at companies including Twitter and Elastic, focusing on what developers actually need: clear documentation, working examples, authentic advocacy grounded in the developer experience, and feedback loops that bring developer insights to the people building the product.
Since 2017, she has been active in NYC’s creative and live-coding scenes, performing live-coded noise music and visuals as Messica Arson, using modular synthesizers and voice samples across North America, Europe, and Asia. She organizes RAID NYC, contributes to LivecodeNYC, and speaks regularly at conferences including DevRelCon, All Things Open, PyCon, Write the Docs, and PyOhio. Working across Python, JavaScript, DevOps, and creative coding, she writes and builds to make complex systems feel understandable, memorable, and useful.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Deploying AI applications often involves containers, GPU configuration, dependency management, and infrastructure scaling. Runpod Flash offers a simpler, code-first approach. It lets you define remote functions and hardware requirements in Python and run them on serverless GPUs.
In this talk, we’ll explore how Flash moves Python workloads from local development to cloud deployment, examine its underlying programming model, and build a GPU-backed endpoint. Along the way, we’ll discuss where Flash fits in the AI development stack, the problems it solves, and the trade-offs developers should consider. Attendees will leave with a practical understanding of how to turn local AI code into a scalable service without needing to become infrastructure experts.
Vicente Rubén Del Pino Ruiz
Vicente Rubén Del Pino Ruiz
About the Speaker:
With 17 years of experience, I specialize in designing, architecting, and implementing scalable AI and data solutions. I’ve delivered transformative initiatives for several Fortune 5 companies, and I now lead Artificial Intelligence and Data Solutions in the healthcare sector at UnitedHealth Group. I also serve as an Independent Advisor on the European Commission’s Horizon Europe program.
I have extensive experience across cloud-based AI and data environments on Azure and AWS, and in recent years have worked with large language models from OpenAI and Anthropic, delivering AI-driven solutions that serve millions of customers.
I am a regular speaker at industry conferences including the World Data Summit, TDWI, DataFest, Graph+AI Summit, PyData Global, and Airflow Summit, and have been featured in Forbes for my work in data engineering. I also write about LLM architectures, scalable AI, and data innovation on my Substack, “”Sailing the Data Ocean.”””
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Single-shot evaluation, the shape every current AI eval framework ships, is structurally blind to the failures that take agents down in production: patience that runs out at turn six, users who abandon silently, tool calls that lose conversation state, partial successes that masquerade as wins. This talk argues for a different shape, drawn from the digital twin literature and how the approach is already applied in aerospace, autonomous vehicles, and civil engineering: simulate the population of users your agent will meet, run them against the agent, watch what breaks before any human sees it.
Three take aways:
- Why current AI agent evaluation cannot detect the failures that hit production. The structural reason prompt-and-grade testing (the shape every current framework ships) is blind to patience, abandonment, conversation state, and partial success.
- What a realistic synthetic user population looks like, drawn from the digital twin literature in other engineering industries. Four properties: continuous trait clusters instead of fixed persona archetypes; probabilistic dropout (real users walk away silently rather than say goodbye); three-way goal scoring (each conversation is scored “achieved,” “partially achieved,” or “not achieved,” instead of just pass/fail); and variable conversation length (each conversation ends when the persona’s patience runs out, not at a fixed turn count).
- How to wire population-scale evaluation into a release gate. Why run-over-run delta on a held-constant population is the regression-detection primitive AI agents actually need, and how to treat it as a governance lever rather than a research experiment.”
Shamindra Peiris
Shamindra Peiris
About the Speaker:
Shamindra Peiris, alumnus of University of Colombo has 13+ years of experience and expertise in Machine Learning, Analytics and AI based Product Management across multiple industries and sectors ranging from Banking & Finance, Fintech, Agriculture and Manufacturing in North America and in South Asia.
Currently she plays a key role in the Global Predictive Fraud Intelligence team at Visa as a Senior AI Product Manager, driving and leading the design and execution of Real Time payment Risk Analytics product suite.
Further, she serves as an Adjunct Professor and a Program Advisory Committee Member for multiple colleges in North America and volunteers as a mentor for several Non-Profit organizations to uplift and guide young professionals and new immigrants in North America.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This talk will discuss how to leverage generative AI techniques to enhance fraud detection models. Main challenge faced by Fraud Detection models is the class imbalance where legitimate transactions vastly outnumber fraudulent ones.
Creating synthetic data will accommodate robust testing and model training. Further, synthetic data allows testing to fine tune models in a risk-free environment complying to Data Governance while minimizing Data Privacy concerns.
Secondly, this talk will discuss how to incorporate generative AI into the machine learning strategy to introduce novel fraud detection methods and improve predictive accuracy.
Integrating generative AI into ML strategy will shift the focus from static, rule-based detection to generating behavioral intelligence. Unlike traditional systems, generative models learn and adapt in real-time, evolving to recognize new fraud types as they emerge without requiring manual model updates. Further, it will contribute in enhancing accuracy of the model while limiting the False Positives.
At the same time, Generative AI techniques can be used to analyze unstructured data such as texts, notes etc rather than limiting to structured data points.
Lastly, this talk will discuss how to utilize generative AI to simulate potential fraud scenarios, aiding in the development and refinement of fraud prevention solutions.
Nalin Dadhich
Nalin Dadhich
About the Speaker:
Nalin Dadhich is a Sr Machine Learning Engineer working at NVIDIA. He has 7+ years of experience building scalable ML solutions all the way from conventional CV models to LLMs, VLMs and World foundational models like Cosmos. He was the core author of Cosmos 3 WFM released by NVIDIA
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Real-world video just doesn’t capture the moments that matter most for safety – a robot’s grip slipping mid-task, a car cutting it close at an intersection, a forklift drifting too near a warehouse worker. These things are rare by nature, and far too risky to stage or collect at any real scale. So physical AI world models end up blind in exactly the spots where a mistake costs the most – and with nothing to learn from, those blind spots stay hidden until the real world exposes them.
Siddharth Jain
Siddharth Jain
About the Speaker:
Siddharth (Sid) Jain is an AI engineering manager at OpenAI, where he leads work on enterprise AI systems and governed agent workflows. He has more than 10 years of experience across AI, finance, data, and enterprise platforms, with a focus on safe automation, retrieval, auditability, and production controls. He speaks and writes about turning AI prototypes into reliable systems that organizations can trust.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
Teams often give an agent a service credential, add a human approval step, and call the workflow governed. That design breaks down when the agent can revise payloads, retry writes, chain tools, or act across systems with different permission models. The result is an accountability gap: the organization can see that a service account acted, but not necessarily who authorized the business intent, which payload was approved, whether a retry duplicated work, or what changed between proposal and execution.
This session presents a production control model for agent workflows that make consequential writes. It shows how to separate the business-intent identity from the concrete operation; classify tools by impact; issue short-lived, least-privilege credentials only after validation; bind human approval to a canonical payload hash and policy version; execute through controlled services with idempotency keys; and reconcile external state before declaring success. It also covers the evidence record needed to answer four operational questions: who requested the action, what the agent proposed, what policy and human approved, and what the downstream system actually did.
The talk focuses on failure modes that appear after the demo works: stale approvals, overbroad agent permissions, duplicate writes after retries, silent policy changes, and actions whose outcome is unknown. Attendees will leave with a lifecycle they can map to their own platform, concrete interfaces between model output and deterministic controls, and an audit schema that supports incident response, compliance review, and day-to-day operations without turning every agent into a bespoke security project.
What You’ll Learn:
- Do not let the model hold broad standing credentials; mint scoped identity for each authorized operation.
- Bind approval to material fields, a canonical payload hash, and the policy version so edits invalidate stale consent.
- Treat every consequential write as a durable state machine with explicit recovery paths, not as a single tool call.
- Separate retry from reconciliation so an ambiguous timeout does not become a duplicate action.
- Design the audit record as an operating interface: it should reconstruct who requested, what was proposed, what was approved, what executed, and what the downstream system recorded.
Prerequisite Knowledge:
Simon Obstbaum
Simon Obstbaum
About the Speaker:
Simon Obstbaum is a software engineering researcher working with Stanford University’s Software Engineering Productivity Research Group (SWEPR), where he studies AI’s impact on engineering productivity, code quality, and team performance. Previously CTO at Crunchyroll, he led distributed engineering teams building video streaming services serving over 100 million users. His work combines empirical research with firsthand experience scaling technology and leading large engineering organizations.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
AI is reshuffling software engineering: widening performance gaps, changing who the top performers are, and shifting the bottleneck from writing code to directing and reviewing it.
Drawing on research from Stanford’s Software Engineering Productivity Research Group, this talk explores why similar teams achieve dramatically different results with AI. We examine how committed project context, structured agent workflows, and codebase quality relate to productivity and maintainability, alongside a case study following AI adoption across approximately 600 engineers.
Attendees will leave with a clearer understanding of why AI productivity gains vary so widely, how structured AI practices and codebase quality influence results, and why faster coding does not automatically translate into better organizational performance.
Speakers Corner
Jenny Morris
Jenny Morris
About the Speaker:
Jenny Morris has spent more than a decade working with search and retrieval systems, first as an enterprise customer and later across Elastic, Weaviate, and Vespa. She works with engineering and executive leaders to design production AI search, RAG, ranking, recommendation, and agentic search systems, translating corpus scale, relevance, latency, freshness, access control, deployment, and cost into durable technology decisions. Her customer- and provider-side experience gives her a practical view of architecture, operations, buying criteria, and market fit. She writes and speaks about multi-stage ranking, RAG at scale, personalization, agent-facing retrieval, and production trade-offs that determine whether AI systems create lasting value.
Date and Time:
Talk Title:
Abstract:
AI agents are only as good as what they retrieve. On the BrowseComp-Plus benchmark, GPT-4.1 answered 93% of questions correctly when handed the right documents, and just 15% when the same model had to find them using standard keyword search (BM25). A single user request can trigger dozens of searches, and each miss shapes what the agent looks for next.
In this session, we will follows an AI shopping assistant as it investigates a product, compares alternatives, and checks availability. Each step needs a different kind of search: exact matching, image similarity, structured filters, or a different ranking strategy. Each result shapes the next step, so one bad retrieval can derail the whole task.
You’ll see why most search interfaces were built for humans, not agents, and how to expose the controls agents actually need. We will also introduce Agent-Ready Schema and Schema-to-Skill, an approach that turns your search schema into guidance agents can use to build their own queries.
Chris Matteson
Chris Matteson
About the Speaker:
Chris Matteson is a technologist with a love of software plumbing and open source. At Union he Heads the Sales Engineering team and helps customers from startups to frontier AI companies and fortune 100 build for the future.
Date and Time:
Talk Title:
Abstract:
As AI continues to revolutionize everything, it’s clear that it’s not enough to build a wrapper around a foundational model. The competitive companies of today and tomorrow will harness their data and institutional knowledge to empower AI factories that provide a durable competitive advantage. Those that don’t will be increasingly eroded by generalist AI solutions.
We will go through what an Sovereign AI Factory is, and why it’s durable in the face of increasingly sophisticated generalist AI.
Matthew Xu
Matthew Xu
About the Speaker:
Matthew Xu is the Co-Founder and CTO of Agentic Fabriq, a Y Combinator-backed startup building secure identity and authorization infrastructure for AI agents. He previously studied at MIT before leaving after his first year to co-found Agentic Fabriq. His work focuses on solving the security and infrastructure challenges of agentic AI, enabling autonomous agents to securely access tools and act on behalf of users.
Date and Time:
Talk Title:
Abstract:
AI agents that work in development often fail in production due to infrastructure issues that basic demos don’t account for: OAuth tokens expiring during background tasks, users revoking access mid-execution, API rate limits shared across thousands of tenants, and large MCP tool catalogs consuming excessive context. This talk examines four common failure scenarios through short live reproductions, followed by a technical walkthrough of the architecture required to address each one. Using Agentic Fabriq Developer, we’ll demonstrate how centralized authentication, credential management, rate-limit handling, and dynamic tool discovery can simplify the infrastructure required to run agents reliably at scale.
Alex Jensen
Alex Jensen
About the Speaker:
Alex Jensen is a product leader at Topogy, where he builds infrastructure for repeatable agent workflows across the AI SDLC. Topogy hooks into coding agents and curates context, so a team’s rules and expertise reach the session at the moment they apply instead of getting lost in a prompt file. Before Topogy, he spent seven years as a software engineer at Google and five years in product management. He cares about making AI agents predictable enough for engineering teams to trust them, and he lives in Boulder, Colorado.
Date and Time:
Talk Title:
Abstract:
Your team has written down how it works with coding agents, and the agent follows it about half the time. This talk shows how to bring the right rule into the session at the moment it applies, with a PR workflow and a demo where a team’s database and security expertise show up unasked.
The Next Wave of AI
Lightning Talks
Workshops
Dr. Amir Feizpour
Dr. Amir Feizpour
About the Speaker:
Dr. Amir Feizpour is Founder, CEO and Chief Scientist of Aggregate Intellect, building a generative business brain for service and science companies, and serves as a fractional Chief AI Officer for several startups. Previously he led NLP product at Royal Bank of Canada and did quantum computing research at the University of Oxford, resulting in publications and patents. He holds a PhD in Physics from the University of Toronto. Amir has built a global community of 5,000+ AI practitioners, advises at MaRS Discovery District, and leads R&D collaborations with McGill, the University of Toronto and Amii.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
An evidence-driven workflow for building production AI agents: testing assumptions with prompts and real users before writing code, mapping context flow before setting agent boundaries, and letting decomposition emerge from evidence rather than imposed architecture. Attendees will leave with a reusable ideation workflow, a setup guide for AI development assistants, and a framework for evidence-based architectural decisions, illustrated through a production case study of an agentic meal-planning system built in a few weeks.
Michael Perry
Michael Perry
About the Speaker:
Software is math. Every class is a theorem. The compiler is the proof. And unit tests check our work. Michael L Perry has built upon the works of mathematicians like Marc Shapiro, Pat Helland, and Leslie Lamport to develop a mathematical system for software development. He is the author of The Art of Immutable Architecture and the Jinaga Immutable runtime. He applies his love of software mathematics at Improving, a trusted partner for AI-assisted application development.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
AI-assisted development tools can now quickly produce working software that once took months to hand craft. But that software often is based on the average of opinions on which the model was trained. It lacks the precision and expertise of the team that maintains it, the specific stack that they selected, and the particular problem domain that it serves. To produce high quality solutions at scale, you need to build a custom software factory. The factory is built from standard features available in almost all AI-assisted development tools:
- Skills
- Commands
- Agents
- Workflows
Your team constructs it to capture the specific architectural decisions and implementation patterns that they have developed for the project. They maintain the factory over time as they use it to build artifacts like documentation, code, and tests. Every project is different. Don’t install a generic template. Learn how to construct a software factory for your current project to accelerate development while improving quality and consistency.
Upal Saha
Upal Saha
About the Speaker:
Upal Saha is the co-founder and CTO of bem, infrastructure for building AI-native products that run critical business processes reliably, at scale. He leads the engineering team that runs AI over millions of documents, images and videos a month for customers in logistics, fleet management, automotive and financial services. Before bem he was an engineering manager at Silo and a technical lead at Yext, and spent time on the investing side at Akido Labs. Based in San Francisco.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This is a build-and-break lab for engineers who already know that a demo is not a system. In the first fifteen minutes every attendee stands up a working extraction pipeline against a real invoice, with no schema authoring, and gets structured JSON back. Then we spend an hour breaking it the way production does, and fixing each break with a pattern that transfers to any stack.
Break one: the wrong document. We feed a bill of lading glued to an invoice into the invoice pipeline and watch it confidently produce garbage. Fix: classify before you extract, and make the graph deterministic even though every step inside it is a model. Break two: the answer that is probably right. Models do not tell you how confident they are, so we compute per-field confidence, find the fields that fall below 95%, and route those, and only those, to a human whose correction feeds back into the system. Break three: the messy string. “10 cases organic gala apples, 88 ct” has to become one SKU; we show why canonicalization is its own step, how to score matches, and where to set the threshold for review. Break four: the hostile input. Everyone runs an image whose pixels read “ignore all previous instructions” and we discuss, with the result on screen, what it means to treat inbound data as data rather than control. We close by labeling a handful of outputs and running a regression test between two versions of the pipeline, because evaluation is a loop, not a phase.
Attendees leave with a running pipeline in their own account and five patterns they can apply on Monday regardless of vendor: classify-then-extract, confidence scoring for models that lack it, threshold-based exception routing with feedback, semantic security checks on inbound data, and versioned regression testing. The platform used in the room is ours; the lessons are not.
Matt Foley
Matt Foley
About the Speaker:
Matt partners with enterprise clients across sales, onboarding, and technical support, helping teams build reliable, AI powered document workflows. His work sits at the intersection of customer strategy and hands on technical problem solving.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
This is a build-and-break lab for engineers who already know that a demo is not a system. In the first fifteen minutes every attendee stands up a working extraction pipeline against a real invoice, with no schema authoring, and gets structured JSON back. Then we spend an hour breaking it the way production does, and fixing each break with a pattern that transfers to any stack.
Break one: the wrong document. We feed a bill of lading glued to an invoice into the invoice pipeline and watch it confidently produce garbage. Fix: classify before you extract, and make the graph deterministic even though every step inside it is a model. Break two: the answer that is probably right. Models do not tell you how confident they are, so we compute per-field confidence, find the fields that fall below 95%, and route those, and only those, to a human whose correction feeds back into the system. Break three: the messy string. “10 cases organic gala apples, 88 ct” has to become one SKU; we show why canonicalization is its own step, how to score matches, and where to set the threshold for review. Break four: the hostile input. Everyone runs an image whose pixels read “ignore all previous instructions” and we discuss, with the result on screen, what it means to treat inbound data as data rather than control. We close by labeling a handful of outputs and running a regression test between two versions of the pipeline, because evaluation is a loop, not a phase.
Attendees leave with a running pipeline in their own account and five patterns they can apply on Monday regardless of vendor: classify-then-extract, confidence scoring for models that lack it, threshold-based exception routing with feedback, semantic security checks on inbound data, and versioned regression testing. The platform used in the room is ours; the lessons are not.
Rajiv Shah
Rajiv Shah
About the Speaker:
Rajiv Shah is the Agentic AI Engineer at OpenHands with a passion and expertise in Practical AI. He focuses on enabling enterprise teams to succeed with AI. Rajiv has worked on GTM teams at leading AI companies, including Hugging Face in open-source AI, Contextual AI in context engineering, Snorkel in data-centric AI, Snowflake in cloud computing, and DataRobot in AutoML. He started his career in data science at State Farm and Caterpillar.
Rajiv is a widely recognized speaker on AI, published over 20 research papers, been cited over 1000 times, and received over 20 patents. His recent work in AI covers topics such as sports analytics, deep learning, and interpretability.
Rajiv holds a PhD in Communications and a Juris Doctor from the University of Illinois at Urbana Champaign. While earning his degrees, he received a fellowship in Digital Government from the John F. Kennedy School of Government at Harvard University. He is well known on social media with his short videos, @rajistics, with over 100k followers.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
You can start a simple agent with a model and a prompt. But you soon realize that improving it requires adjusting what the model can see, what it can do, how it remembers progress, which model handles each step, and what evidence allows the work to stop. Harness engineering is how those pieces become one working system.
This workshop uses coding agents to make that system concrete. Participants will work through six decisions: choosing a harness, designing tools and retrieval, placing context and memory, routing models and reasoning, controlling goals and validation, and deciding when a task benefits from multiple agents. Through hands-on experiments, we will change these settings and inspect what happens in the trace and final result.
Each exercise is paired with an overview of current research on agent harnesses. Together, the research and experiments show how the major components interact and where each approach reaches its limits. Participants will leave with a practical way to reason about the whole agent, not just its model.
Chris Alexiuk
Chris Alexiuk
About the Speaker:
Chris Alexiuk is a deep learning developer advocate at NVIDIA, working on creating technical assets that help developers use the incredible suite of AI tools available at NVIDIA. Chris comes from a machine learning and data science background, and he is obsessed with everything and anything about large language models.
Date and Time:
Talk Title:
Track:
Sub Category/Topic:
Technical Level:
Abstract:
In this workshop we’ll stand up NemoClaw end to end: install the reference stack, get OpenClaw running inside the OpenShell sandbox, configure inference routing, and lock down a network policy that survives a multi-hour agent session. We’ll walk through the blueprint, the CLI, and the approval flow, then run a real long-lived agent against it and break things on purpose so you know what the layers actually catch.
Greg Loughnane
Greg Loughnane
About the Speaker:
Dr. Greg” Loughnane is the Founder & CEO of AI Makerspace, where he builds training programs for engineering and data science teams at mid-market and enterprise companies. Since 2021, he has built and led industry-leading Machine Learning education programs. Previously, he served as an AI product manager, university professor, consultant, advisor, and researcher. He loves trail running, wrenching on cars, and is based in Columbus, Ohio.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
In this workshop, we’ll explore best practices for maximizing ROI (and excitement!) for your retrieval system throughout your entire AI product lifecycle. You’ll understand how to make data and retriever decisions at the MVP, PoC, and Production phases of your project.
We’ll center our workshop around the story of Nadia, a second-line support engineer on the service desk of a regional broadband and mobile operator who is looking to build a retriever for her knowledge base. What should she choose based on the shape of the corpus today, and what kind of corpus change should make her reconsider?
Guided by our heroine, we will explore the key aspects of retrieval; that is, getting the best of both semantic and keyword-based search, using agent loops versus RAG/search pipelines, constructing knowledge bases using embedded chunks versus an LLM Wiki approach, and of course, what happens when Nadia expands the scope of the problem to include multimodal image retrieval.
Scott Askinosie
Scott Askinosie
About the Speaker:
Scott Askinosie, PhD, is a Senior Principal AI Engineer at SteelEngine, where he architects the knowledge and memory layer for governed agentic workflows in regulated enterprises. He built SteelEngine’s three-tier governed memory system (session, team, org) with deterministic policy gates, novelty scoring, provenance, and human endorsement checkpoints. Previously he spent six years as Lead Data Scientist and Professor at WGU, building student retention analytics across 487K+ records, and served as Lead Technical Trainer at Weaviate, delivering 20+ enterprise workshops including NATO, FactSet, and Fannie Mae. Scott holds a PhD in Quantitative Biology from the University of Missouri, leads the 400+ member LangChain Austin community, and speaks at ODSC and Reuters Momentum.
Date and Time:
Talk Title:
Track:
Technical Level:
Abstract:
This is a hands-on build. Attendees start with an ungoverned append-only memory on a vector database, break it deliberately, and then rebuild it as a governed memory with tiered scope.
We work through the write path step by step: capturing a candidate memory, scoring it for novelty against what is already stored, attaching provenance, and passing it through a deterministic, fail-closed gate before it is promoted from session scope to team scope to organization scope. Attendees implement the gate themselves rather than calling a library, because the point of the exercise is understanding exactly which decisions must stay deterministic and which can be model-scored.
The second half is adversarial. Attendees inject false facts into their own memory and measure what survives promotion, then tighten their gates until the poisoning is quarantined. We finish with the evaluation question most teams skip: how to tell “”the agent never received the memory”” apart from “”the agent received it and ignored it,”” and why an eval suite that only scores output quality cannot detect a memory layer that has silently stopped working.
Attendees leave with working code, a reference architecture for tiered governed memory, and a test harness for poisoning resistance they can point at their own stack.
Agenda
We will release it soon.
This agenda is still subject to changes.
Join free virtual session November 16, 2026 then meet us in Austin for in-person case studies, workshops, and expo November 17-18, 2026
Loading...