Human insights driven latent space for different driving perspectives: A unified encoder for efficient multi-task inference
icra 2026 · a. bairouk, m. maras, w. xiao, t.h. wang, p. chareyre, et al.● 7 SHIPPED · 3 LIVE
Selected work
A working archive of vision systems, agent architectures, and the messy bits in between. Most ship to production; a few are research demos.
The robot sees, listens, acts
ODIN humanoid (Capgemini AI Robotics Lab) — a production real-time multimodal stack: voice interface, GPT reasoning, and action execution, with 2D segmentation + 6D object pose and RAG-grounded responses. Shipped for Orano and Carrefour.
GPT APIRAGYOLO6D poseKafkaDocker
A bot that reasons before it acts
A multi-user Telegram assistant built as a uv monorepo — a custom ReAct loop calls calendar, weather and notes through three standalone MCP servers, with Whisper voice input and Qdrant long-term memory.
MCPReActpydantic-aiQdrantWhisperDocker
A harness for coding agents
A from-scratch CLI agent harness over pydantic-ai and OpenAI — a streaming reason–act–observe loop with approval-gated tools, isolated subagents, auto-compression, loadable skills and MCP servers.
pydantic-aiOpenAIMCPReActPythonuv
One encoder. Every view of the road
A single Swin-based encoder trained jointly on depth, pose, 3D scene flow and four kinds of segmentation — frozen, its latent space steers better than ImageNet pretraining. First-author paper with MIT CSAIL, accepted to ICRA 2026.
Swinmulti-taskdepthsegmentationPyTorchMIT CSAIL
Chatbots that read your docs
A live multi-tenant RAG platform — upload your own documents and chat with a bot grounded in them. Next.js on Vercel talks to a Cloudflare Worker for retrieval (Vectorize + D1), with OpenAI embeddings and gpt-5.1, streamed token by token.
RAGNext.jsCloudflareLangChainOpenAIVectorize
Reading disease in a single brain scan
My PhD — deep learning that detects neurological disease from one structural brain MRI. Two interpretable biomarkers, deep grading and brain structure ages, scale from Alzheimer's to six conditions, and ship as four registered tools on the volBrain platform.
deep learningMRIAssemblyNet3D U-NetPyTorchvolBrain
Tracing a gesture from a single chip
My engineering final-year project at CEA LIST — reconstruct the 3D path of a repeated hand gesture from a coin-sized inertial sensor (IMU). A CNN + BiLSTM learns the motion, an error-state Kalman filter grounds it, and a hand-built mirror-and-infrared rig supplies ground truth.
deep learningIMUCNN+BiLSTMKalman filterRaspberry Piquaternions
PUBLICATIONS
Things I wrote down
Peer-reviewed journal and conference work across medical imaging and autonomous-driving perception. Full list on Google Scholar.
Enhancing Steering Estimation with Semantic-Aware GNNs
arxiv preprint · f. makiyeh, p. chareyre, r. hasani, m. blanchon, d. rusBrain structure ages — a new biomarker for multi-disease classification
human brain mapping · m. clément, b. mansencal, p. coupéDeep learning disconnectomes to accelerate and improve long-term predictions for post-stroke symptoms
brain communications · a. matsulevits, p. coupé, l. talozzi, et al.Deep grading for MRI-based differential diagnosis of Alzheimer's disease and Frontotemporal dementia
artificial intelligence in medicine · m. clément, v. planche, b. mansencal, p. coupéTowards better interpretable and generalizable AD detection using collective artificial intelligence
computerized medical imaging and graphics · m. clément, b. mansencal, p. coupé