Technical notes
Architecture decisions from production AI work, written out at the level of detail I would use to answer a follow-up question.
11 August 2026
Architecture of a multi-step AI assistant
An LLM classifier routes intent across 13 skills, context fetches go out speculatively and in parallel, and the answer streams over WebSocket. What that buys, what it costs, and what breaks.
11 August 2026
Testing a non-deterministic system
Assertions stop working once the same input can produce a different answer. Deterministic checks at the bottom, an LLM-as-judge harness above them, human review on a sample, and two production metrics that tell you when it slipped.