[01]
Published 2026-09-04 · 4 min read
Most changes you make to an AI system produce no measurable improvement, and without a gate that can revert them you will never find out which ones. Build the measurement layer first.
[02]
Published 2026-09-04 · 4 min read
An automated writing pipeline can run perfectly for months while reaching nobody. The health metric it reports is its own activity, and the number that matters takes a separate query.
[03]
Published 2026-05-14 · 5 min read
Most public AI businesses fall into one of three categories. Two are crowded. The third is where I work — local-first deterministic AI where the model proposes and code disposes.