About
I'm Tom Vaucourt, an AI engineer. My work keeps tracing the same loop: build the end-to-end thing, hit the layer underneath that's too slow or broken, fix it there, and send the fix upstream. Building localvoxtral led to contributing the Voxtral streaming path in mlx-audio-swift; running coding agents against local models led to a merged llama.cpp fix that benefits anyone pointing Claude Code at their own hardware.
Day to day I run a fleet of coding agents in isolated git worktrees, coordinated from a single session. The tooling that makes this workable is public: a repo-exploration subagent, provider plugins, delegation skills. Tools you keep to yourself don't get better.
This site is where I write about that work and argue for that way of working: strong opinions, with receipts. The full project list lives on GitHub.