Physicist (PhD, ESPCI Paris) building ML systems at scale.
I build inference engines from scratch through Idle Intelligence, my independent lab: LLM and speech models in Rust, quantized, running entirely in the browser on WebGPU and WebAssembly.
By day, I'm the architect of Sonos AI: real-time voice and LLM inference, from the device to multi-region cloud.
Previously AI Tech Lead at Scaleway (inference infrastructure) and co-founder & CTO of Sense AI (physics-informed sensing).
Idle Intelligence · models and data on Hugging Face
Independent work: time-series forecasting and browser-native inference for speech, TTS and LLMs, all client-side.
- lean : my LLM inference engine, one Rust + WGSL codebase that runs natively and in the browser, on WebGPU, CPU threads or a single thread depending on the device; Qwen2, Qwen3 and Llama models, token-exact against HF transformers
- t0-web : from-scratch browser inference engine for time-series foundation models (Rust + WebGPU); matched a published forecasting benchmark at roughly a seventh of the size
- stt-web : real-time speech-to-text, 100% in the browser (Rust + WASM + WebGPU)
- tts-web : text-to-speech, client-side
- ridgeline / astres : WebGPU 3-D visualization of planetary data