Learning inference engineering in public, one day at a time. Every entry links to something you can open and check — a commit, a pull request, a benchmark, a post.
No proof, no entry.
One repo for every measurement rather than one per artifact, so results accumulate somewhere. The rules are the actual content right now: derive the number before measuring it, report the gap rather than the win, disclose the whole config, and publish where an optimisation stops paying for itself. Easy to write on day one and much harder to adopt later.
Both used to lead with job titles. They now lead with numbers I can defend: a platform at ~30K daily visitors, an offline-first POS across 11 terminals, and multi-minute report queries in a 15-year-old database traced to missing keys and no indexing, then cut to under 3 seconds. Also shipped this log — the proofs field is a non-empty tuple, so an entry without evidence fails the build rather than relying on my discipline.