Essays and playbooks on local-first AI, RAG, LLM evals, and systems engineering.
I published a post-mortem naming the exact question I failed to ask, then entered another competition the same day and failed to ask it again. What actually wins these competitions, and why I keep missing it.
What I learned competing in an open ML-optimization competition where the ranked machine was an M5 Max and mine was not - the oracle I found, the levers I priced dead, and why measurement bandwidth decided everything
I built willitcall, a conformance suite and public matrix for tool calling on local models. The findings that survived replication, the ones that did not, and the bug report I had to retract.
A standalone kit that installs a reliable orchestrator-delegates-to-executor workflow into Claude Code - and the dogfooding loop that wrote it
An independent, reproducible evaluation of TabFM - what held up, what did not, and the bug fix that got merged into Google's repo
TurboQuant on an M1 Pro - five bugs across two codebases, one coupled math fix, and 100% needle retrieval at 16K with 3.5x less KV memory