The interesting part of AI engineering is everything underneath the model call.

I build with a fleet of coding agents, and when a layer under them breaks or falls short, I fix it upstream.

Now: on-device speech for Apple Silicon, shipping localvoxtral and contributing the Voxtral realtime inference path to mlx-audio-swift.

GitHub·LinkedIn·Hugging Face·npm

Selected work

  • Native macOS menu-bar app for realtime, fully local dictation: words appear while you’re still speaking, with optional on-device LLM polishing. Runs Mistral’s Voxtral Mini 4B Realtime on Apple Silicon.

  • Contributor to the Voxtral realtime streaming path: incremental mel/conv front end (O(N²) → O(N) per utterance), fp16 dtype fixes, Metal buffer-pool policy. Streaming went from ~1.9× slower than realtime to a flat 0.76 RTF.

  • Merged upstream: preserve Anthropic thinking blocks through the server’s message conversion (#20120).

  • Read-only repository-exploration agent that coding agents delegate to over bash: cited file:line answers, not context noise.

  • voxmlxPython

    Realtime Voxtral transcription in MLX; my fork adds an OpenAI-Realtime-compatible WebSocket server and Metal memory caps, with 4-bit conversions on Hugging Face.

  • Diagnosed a V8 segfault on small-ICU Node builds, filed it upstream with a proposed patch, and shipped a pure-JS Intl.Segmenter fallback to npm the same day.

Writing

All posts →