AI Summit by Unifold
About event
19 November
8 00 am – 5 00 pm
San Francisco, exact venue to follow
50 members only
Engineering and AI leaders (CTOs, Chief AI Officers, VP Engineering, Heads of AI, ML & Platform) shipping AI in production
Outcome: Leave with de-risked answers to the hard calls of shipping AI in production, evals over vibes, agent reliability, inference economics, and securing what agents can actually do.
The premise
Anyone can build something that dazzles in a demo and works three steps deep on a clean input. Almost nobody can tell you what theirs does on step nine with a messy one. The interesting questions are about reliability, cost, security, and what you do when a probabilistic component sits inside a system that has to be right.
This is a builders' room with fifty people who own that Tuesday. Postmortems over pitches, war stories over roadmaps, the calls you actually had to make. What's said in the room stays in the room.
What the day gets into
Evals over vibes. Testing AI before it ships, error analysis, eval-driven development, and why the demo passes when production fails.
The reliability math of agents. Why a 95%-per-step agent fails at step nine, and the architecture that survives real inputs.
Running a probabilistic system. Debugging something that answers differently every time; tracing, guardrails, and catching the confident wrong answer.
The economics of inference. Cost that scales with usage, latency budgets, and the levers, caching, routing, small models, self-host — before the bill arrives.
Model and architecture calls. RAG vs long context vs fine-tuning; and what to do when a silent model update breaks what you shipped.
AI-written code, honestly. Where it helped, where it broke, what it did to review and to the junior-engineer pipeline past the 10x claims.
Securing what agents can do. The lethal trifecta, the expanding attack surface, and giving each agent only the keys it needs.