Bauke Brenninkmeijer

Companion repo for the PyData Amsterdam 2026 talk Building the evaluation flywheel: how human judgment becomes an evaluator you can run on every change. A data-analysis agent answers 50 Sphere.com business cases, a jury of LLM judges flags where it disagrees, and a human's answers to those boundary questions go into a pass/fail evaluator.

A wall of mail slots for coding agents running in parallel. Each git worktree gets a pigeonhole; agents post what they are working on and leave each other notes. No server, no daemon, no MCP. One POSIX shell script and the filesystem.

My personal blog about topics I find interesting in machine learning, deep learning and AI.

Slides and worked example for a talk on evaluating LLM agents: a data-analysis agent over DuckDB, its case corpus and traces, and the jury runs that turn human judgement into an aligned LLM judge.

Security of AI Agents, AI Engineer 2026 edition. Forked from the ADC talk repo: slide deck plus a live goal-hijacking red-team campaign against an orq.ai agent.

Companion repo for the ADC 2026 talk "Security of AI Agents": the slide deck, plus a live goal-hijacking red-team campaign against Clarabelle, an orq.ai agent whose only goal is to be a cow.