Real-time conversational system
Designed and deployed a production-grade conversational system on GPT-based APIs — the robot answers questions, reacts, and drives its own responses from live dialogue rather than a fixed script.
★ FEATURED · CAPGEMINI AI ROBOTICS & EXPERIENCES LAB● LIVE
The robot sees, listens, acts
Humanoid robotics — real-time multimodal interaction & perception for ODIN, deployed with Orano and Carrefour.
- client
- Orano + Carrefour
- role
- Applied AI Engineer
- period
- 2024 → now
- platform
- ODIN humanoid · 1.35 m
- location
- Paris, FR
perception
interaction
01OVERVIEW
ODIN — Observe, Detect, Interact, Navigate — is the humanoid built at Capgemini's AI Robotics & Experiences Lab. I own the multimodal interaction and perception stack that lets it hold a genuine conversation, understand the scene in front of it, and act on what it's asked.
It ships to real clients: Orano, where the robot takes on maintenance in environments that are risky for people, and Carrefour in retail. The hard part was never the demo — it was making perception and dialogue reliable enough that a non-expert can rely on the robot in a real room, in real time.
02WHAT I BUILT
Multimodal interaction pipeline
Combined a voice interface, LLM reasoning, and action execution into a single loop so that speech, understanding, and movement stay in sync.
Real-time visual perception
Delivered 2D segmentation and 6D object pose estimation in real environments — the robot identifies people and objects and localizes them in 3D to act on them.
RAG grounding
Integrated retrieval components so responses are grounded in project and environment context instead of the model's priors alone.
03STACK
AI · Vision
GPT APIRAGOpenCVYOLO2D segmentation6D pose
Systems
PythonDockerKafkamicroservices
04REFERENCES