The engineering decisions behind production systems — the tradeoffs, not the buzzwords.
20 Aug 2026
Decisions lived in chats, then in two docs folders, then in one knowledge base that Claude Code searches by meaning. The episode that convinced me, and the rule that came with it: a found note is a claim, not a fact.
26 Jul 2026
I had four models audit the same Prisma schema twice each, verified every claim by script, and still implemented a critical finding the application had already solved. What eight AI reviews are good for, and what they are not.
05 Jul 2026
More than 150 models, legacy pages still in production, and traffic data I could not trust. How I worked out which models a limited set of functions actually depends on, where the script needed a human, and what the result does not prove.
15 Jun 2026
A self-hosted AI agent with local embeddings, hybrid retrieval, and persistent memory over my own knowledge base — and why the hard part wasn't the model.
20 May 2026
My first mobile app started at 700MB, climbed past 1.2GB, and got OOM-killed on real devices. The full teardown of how it reached a stable 350MB — and why the real leak was the shape of the app, not a setting.