6 posts
Writing
Reproductions, post-mortems, and debugging stories. Every claim carries the measurement behind it, including the ones I got wrong.
- The Work You Don't Do: Losing Two Optimization Competitions the Same WayI named the question I failed to ask, then entered another competition and failed to ask it again. What actually wins these races, and why I keep missing it.2 competitions entered2 lost the same way
- Eight Submissions, Zero Promotions: A Week Inside mlx.fast on the Wrong HardwareCompeting in an open ML-optimization competition on hardware that could not run the ranked benchmark, and why measurement bandwidth decided everything.8 submitted0 promoted
- Same Weights, Same Prompt, Opposite Results - Why Local Tool Calling BreaksI built willitcall, a conformance suite and public matrix for tool calling on local models. The findings that survived replication, and the one I retracted.Blamed the modelit was the stack, finding retracted
- Baton: I Built a Tool for Delegating Code to an AI, Then Used It to Build ItselfA standalone kit that installs a reliable orchestrator-delegates-to-executor workflow into Claude Code - and the dogfooding loop that wrote it8 tasks delegated2 failed, both my spec's fault
- I Tried to Break Google's New Tabular Foundation Model. Then I Fixed It.An independent, reproducible evaluation of TabFM - what held up, what did not, and the bug fix that got merged into Google's repoWon on 10 datasetsdemoted 2 of my own wins to ties
- From 0% to 100%: Debugging a KV Cache Compression Algorithm on a 16GB MacBookTurboQuant on an M1 Pro - five bugs across two codebases, one coupled math fix, and 100% needle retrieval at 16K with 3.5x less KV memoryCalled it a bug fixvariance reduction, corrected