AI engineer — Local inference and agentic developer tooling. I build end-to-end systems and contribute upstream to the tools I depend on.

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

• localvoxtral — 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, or against any OpenAI Realtime-compatible backend. Swift
• mlx-audio-swift — 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, and quantized tied-embedding loading — teaching the engine to load a 4-bit tied-head checkpoint I publish, which cuts the per-token LM-head projection ~10× (~30 ms → ~3 ms on an M1 Pro). Streaming went from ~1.9× slower than realtime to a flat 0.76 RTF. Swift
• llama.cpp #20120 — merged: preserve Anthropic thinking blocks through the server's message conversion. C++
• fastcontext — read-only repository-exploration agent that coding agents delegate to over bash: cited file:line answers, not context noise. Python
• voxmlx — realtime Voxtral transcription in MLX; my fork adds an OpenAI-Realtime-compatible WebSocket server and Metal memory caps. 4-bit conversions on Hugging Face. Python
• pi-intl-segmenter-fallback — diagnosed a V8 segfault (null icu::BreakIterator on small-ICU Node builds, nodejs/node#51752), filed it upstream with a proposed patch, and shipped this pure-JS Intl.Segmenter fallback to npm the same day. TypeScript
• pi-llamacpp-provider — pi extension that registers a llama.cpp server as a provider and auto-discovers every model it exposes. JavaScript
• skills — skills that let a coding agent delegate work to another coding CLI and keep itself in sync from git. Shell

• Plugin userConfig options are not expanded in http hook headers (${CLAUDE_PLUGIN_OPTION_*} resolves empty) anthropics/claude-code
• Hosting: llama.cpp needs --reasoning-format none microsoft/fara
• Recovered daemon-transport bounce leaves a permanent error on the workspace row manaflow-ai/cmux
• VoxtralRealtime: load checkpoints with a quantized tied embedding Blaizzy/mlx-audio-swift
• Hoist per-layer-invariant attention inputs out of the Voxtral layer loops Blaizzy/mlx-audio-swift
• VoxtralRealtime streaming: stop clearing the Metal buffer pool on every step Blaizzy/mlx-audio-swift
• VoxtralRealtime streaming: incremental mel/conv front end (O(N²) → O(N) per utterance) Blaizzy/mlx-audio-swift
• Fix float32 leak in VoxtralRealtime streaming Blaizzy/mlx-audio-swift

LinkedIn · Hugging Face · npm