Staff AI Engineer building at the intersection of AI, robotics, and edge computing.
Based in Paris. Currently at Air Liquide. Focused on making AI systems that run locally, respect privacy, and work on consumer hardware.
lerobot-jetson — Docker image bringing CUDA-accelerated PyTorch to LeRobot on NVIDIA Jetson Orin. No official channel ships a cp312 CUDA torch wheel for Jetson, and the usual community fallback (jetson-containers) has been unmaintained since 2025 — this builds torch, torchcodec, and torchvision from source instead. Verified on real Orin hardware: GPU-accelerated video decode, working torch.distributed, real policy inference. Full build ~5.5h, 92% of that compiling PyTorch alone. Mirrored into huggingface/lerobot's own repo as a community-maintained Dockerfile (#4599).
Baby Reachy-Mini Companion — An autonomous AI nursery companion built on the Reachy Mini robot. 7 AI models (VAD, STT, TTS, YAMNet, YOLO, LLM, VLM) orchestrated on-device with zero cloud dependency. The robot listens, sees, reasons via tool-calling SLMs, detects baby cries, spots dangerous objects, and alerts parents — all running locally on a Mac or a $700 NVIDIA Jetson Orin NX.
- Fully local pipeline: Silero VAD → Faster-Whisper → Ollama/vLLM (3B-4B SLMs with tool calling) → Kokoro TTS
- Autonomous safety: YAMNet cry detection + YOLO danger scanning with deterministic alerts (never gated on LLM behavior)
- Deployed on NVIDIA Jetson Orin NX via vLLM and llama.cpp with quantized models (~3s end-to-end latency)
- Built for the NVIDIA GTC x Hugging Face Golden Ticket Contest — ranked #2 in community rankings
lerobot-policy-turbovla-so101 — Porting TurboVLA (a lightweight vision-language-action model — DINOv3 + BERT, no full VLM backbone) into LeRobot and validating it on a real SO-101 arm, not just simulation benchmarks. Found and diagnosed a real training failure along the way: a checkpoint that looked healthy by every loss-curve signal had actually learned a proprioceptive shortcut (predicting action ≈ current state, nearly ignoring both camera input and language instructions) — root-caused, fixed by warm-starting from the original paper's fine-tuned vision weights and unfreezing the backbone, then re-verified with an input-sensitivity gate.
MCPH — MCP Hub (mcph.io). AI artifact storage and sharing platform built on the Model Context Protocol. Upload, organize, and share AI-generated content across Claude, ChatGPT, and any MCP-compatible tool. Features semantic search, poll collection, embeddable badges, and a full MCP server. TypeScript / Next.js / Firebase.
EvalNow — A lightweight AI evaluation tool (evalnow.xyz). Upload a dataset (CSV/JSONL/Excel), get instant AI-powered scoring with detailed reasoning. Privacy-first: no data storage, in-memory processing only. Python / FastAPI.
- huggingface/lerobot —
ChunkSafetyProcessorStep, a policy-agnostic guard against corrupted/out-of-bounds predicted action chunks (#4241); fixed a future-frame information leak in VLA-JEPA's world-model training objective (#4381); Jetson Dockerfile + install docs, mirrored fromlerobot-jetson(#4599); Reachy Mini teleoperation/motor-control integration; co-authored fixes to thedraccusCLI-config error surface (#4260, #4265) - langchain-ai/langchain-aws — Added
llms.txtdocumentation for AI coding tool support
I write about deploying AI in production on Medium — covering RAG pipelines, LLM serving on AWS SageMaker, edge deployment, and agentic architectures.
- Agentic AI — Tool-calling agents, multi-step reasoning, MCP integrations, autonomous decision loops
- Local/edge AI — Running multi-model pipelines on consumer hardware without cloud dependencies
- Robotics — Autonomous companion robots, assistive robotics, teleoperation, LeRobot ecosystem
- SLM engineering — Tool calling with 3B-4B models, working around reliability limits, quantization for edge deployment
- NVIDIA Jetson — vLLM / llama.cpp deployment, memory optimization, DLA core utilization
Python TypeScript PyTorch ONNX Ollama vLLM llama.cpp FastAPI Next.js YOLO Whisper LangChain AWS GCP Docker NVIDIA Jetson