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

Now: Starbridge, so a coding agent that stops for a question reaches you on your phone, live at starbridge.run. Still shipping voice for coding agents in localvoxtral.

• Starbridge — when a coding agent (Claude Code, Codex, Pi, opencode) stops for a question, Starbridge puts it on your phone and in your browser. You answer with one tap and the session carries on. It also shows what's left on each AI plan, read from CodexBar. The server stores your content only as ciphertext, though it still sees who sent what, and when. Free server at starbridge.run, or host your own. TypeScript Kotlin
• vidtheque — the talks you don't have time to watch, queryable by an agent over MCP: transcripts, on-screen text and keyframes, every answer stamped with the second it happened. Self-hosted, with a demo that indexes 310 AI Engineer talks and both days of AI Engineer Paris 2026. Python JavaScript
• localvoxtral — native macOS menu-bar app for realtime, fully local dictation. Words appear while you're still speaking. Built for prompting coding agents by voice, it grounds LLM polishing in the session under your cursor, its screen and the repo's vocabulary. Swift
• mlx-audio-swift — on the Voxtral realtime streaming path: incremental mel/conv front end (O(N²) → O(N)), in-place KV append and token-by-token decode for bounded memory, and the 4-bit tied-head checkpoint I publish. Streaming went from ~1.9× slower than realtime to 0.76 RTF. Since October, fixes across its other speech models too. Swift
• working-set — how many agents a given GPU configuration keeps warm, and which constraint binds first: KV cache, decode bandwidth, or prefill compute. A scenario model on PyPI (ws predict, ws test on a live vLLM endpoint) behind an explorer that answers with a verdict. Python HTML
• llama.cpp #20120 — merged: preserve Anthropic thinking blocks through the server's message conversion. C++
• skills — the agent skills I run across Claude Code, Codex, opencode and pi: delegating work to another CLI, Remote Control servers under systemd, prose cleanup. Each one is a gotcha backed by evidence, deleted once upstream fixes it. Shell

• server : realtime: reject odd-length audio appends ggml-org/llama.cpp
• Qwen3-ASR: default to 60 s chunks instead of 1200 s Neywa-Labs/mlx-audio-swift
• Nemotron: honor transcription_delay_ms in streaming sessions Neywa-Labs/mlx-audio
• Misc. bug: server: realtime session.update asking for 24 kHz is accepted, the audio is still read as 16 kHz ggml-org/llama.cpp
• server : reject a realtime session.update with another sample rate ggml-org/llama.cpp
• server : report stop_sequence as stop_reason in /v1/messages ggml-org/llama.cpp
• NemotronASR: add Nemotron Unified Neywa-Labs/mlx-audio-swift
• feat(turn-priority-fairness): enable tool calling on the GPU vLLM overlay llm-d/llm-d

LinkedIn · Hugging Face · npm · PyPI