14 of 14 sessions
Glenn McmillanStaff Engineer, Parker LLCGlenn on what separates demos from dependable agent products. The gap is rarely the model. It is evaluation, error handling, and the unglamorous operational work that turns a promising prototype into something a team can put in front of paying customers. Drawing on two years of shipping agents into production, this keynote lays out what to build first and what can safely wait.
Latasha WaltonDistinguished Engineer, Stephens Scott and FinleyMost agent demos work once and fall apart on the second run. This talk walks through the loop itself: how to bound retries, when to let a tool failure propagate instead of swallowing it, and how to keep the working state small enough that a model can still reason about it. Includes failure traces from three production agents and what each one taught us about where loops actually break.
Jessica PollardChief Architect, Beard PLCSoftware that users can reshape at runtime asks different questions of the architecture. In this hands-on session you build a small app whose behavior an agent can rewrite, then break it in the ways real users will. Expect to spend most of the time in code rather than slides. Bring a laptop with Node 20 or later installed.
Michael LucasIndependent Consultant, Mitchell-LeeAn eval suite that only runs before launch tells you nothing about the model you shipped. We cover building a graded set from real traffic, choosing metrics that move when quality moves, and wiring the whole thing into the deploy so a bad change never reaches a user. Bring a product with an eval gap and leave with a plan for closing it.
Cynthia YoungSenior ML Engineer, Reynolds Ayala and WilliamsSwapping the model underneath a live product is a migration, not a config change. This session covers shadow traffic, per-prompt scorecards, and the rollback path you need in place before the swap, plus the class of regressions that only show up under real user phrasing. We walk through two swaps that went fine and one that had to be reverted at 2am.
David CainHead of Applied AI, Moss Peterson and RossDavid shows how Moss Peterson and Ross wires evals into every deploy. Each prompt change opens a pull request, runs against a graded set, and reports a score diff a reviewer can actually read. The talk covers how that set was built, how it is kept current as the product changes, and what the team does when a score drops but the change is still right.
Nicholas ShortCo-founder & CTO, Jordan Holmes and GomezNicholas's field guide to retrieval that survives production traffic. Covers chunking that respects document structure, hybrid search for the queries embeddings alone keep missing, and the caching layer that kept p99 flat while the corpus grew ten times. Includes the cost model that made the case for a smaller index and a bigger reranker.
Lunch is served in the main hall. Vegetarian, vegan, and gluten-free options are labeled at each station.
Juan SosaPlatform Lead, Smith Dixon and LutzJuan on streaming, speculative tool calls, and where the milliseconds go. A multimodal agent spends its time in places a text-only one does not, and the usual profiling advice misses most of them. This talk gives you a budget you can hold each component to, and shows the three changes that took one agent from four seconds to under one.
Kenneth EnglishResearch Engineer, Gomez GroupKenneth's hands-on session. You build a small fleet of coding agents that split a task, work in isolated worktrees, and merge without stepping on each other. We cover task decomposition, the review step that catches a bad agent before its work lands, and how to keep the whole thing debuggable when it goes wrong. Bring a laptop with git and Node installed.
Lauren TaylorEngineering Manager, Williams Brown and GreerLauren on catching silent quality drops within hours of a model swap. The talk covers the sampling strategy that makes a small graded set representative, an alerting threshold that does not fire on noise, and the triage path once it does. Includes a walkthrough of a regression found, diagnosed, and reverted in a single afternoon.
John BrayFounding Engineer, Moore-NewtonJohn's war stories from Moore-Newton's platform team. Reaching 40,000 queries per second meant giving up on a single index, batching aggressively at the edge, and accepting a staleness window the product team had to sign off on. This talk covers the architecture, the two rewrites it took to get there, and the monitoring that keeps it honest.
Building an event copilot live on stage with tool calls against a real schedule. Starting from an empty file, we wire a model to the conference API, give it the four tools it needs, and watch it answer attendee questions in real time. Nothing is pre-recorded, so expect at least one thing to break and get fixed in front of you.
Sean LopezVP of Engineering, Hudson-JohnsonGlenn McmillanStaff Engineer, Parker LLCDavid CainHead of Applied AI, Moss Peterson and RossSean moderates speakers from across the program. The panel takes questions on the hardest part of the last two years: deciding when a prototype is ready, what to cut to get it there, and how a team stays confident once the model underneath it keeps changing. Submit questions during the day and the best ones go to the panel.