Talks

Four talks I can give, each built from systems I've run and mistakes I've made, not from slides about the future. Each one works as a 30–40 minute talk or a hands-on workshop.

AI writes code. You still own the failure.

Three real bugs AI-assisted code handed me: a hook after an early return, chunks that silently wiped an index, a reranker that died with an expired key. None of them were caught by reading the diff. The talk covers the gates that did catch them (lint rules, evaluation sets, circuit breakers, traces) and how to decide what to hand to a model and what to keep.

• For: engineering teams adopting AI-assisted development
• You leave with: a checklist of gates that catch what review misses

RAG that knows when to shut up

A confident wrong answer does more damage than "I don't have that". Per-mode relevance floors, browse intent versus question intent, structural-node guards, and a fallback answer engine that works with no model at all, shown on a live assistant with its failure cases.

• For: developers building retrieval and assistants
• You leave with: patterns for abstaining, degrading and citing honestly

From 22 AI buttons to one useful decision

AI tools love to expose every capability as a button. This talk is a before/after of an AI editor panel: why equal-weight actions fail, how to recommend one next step from real state, and how to write labels around outcomes instead of tools.

• For: product designers and engineers building AI features
• You leave with: a pattern for turning an AI toolbox into guidance

Building an AI system you can actually measure

I benchmarked my own retrieval default and it lost. A hand-labeled set, an ablation across five configurations, answer-level evals with the model on and off, latency and cost from traces, and what to do when the numbers say your architecture decision was wrong.

• For: teams putting RAG or assistants into production
• You leave with: a minimum viable evaluation setup you can build in a day