Dec 2025 – Present · 4 mos · Paris, Île-de-France, France · Hybrid

• I founded 20n Research Laboratory to solve the fundamental incompatibility between current AI and the physical world. While LLMs consume megawatts to process text in the cloud, biological brains run consciousness and complex motor control on just 20 watts. My goal is to bridge this efficiency gap of one million times to unlock true embodied autonomy.
• The laboratory focuses on developing a universal neural architecture capable of controlling any robotic body with extreme energy efficiency and zero reliance on cloud infrastructure. We are moving away from the frozen weights of traditional deep learning toward continuous-time Liquid Neural Networks and Active Inference models. This approach allows robots to adapt to new morphologies or environments in seconds, not months, by minimizing surprise rather than maximizing arbitrary rewards.
• We are currently benchmarking these architectures on MuJoCo physics simulations with a focus on sim-to-real transfer for quadruped and eventually humanoid platforms. 20n is not building chatbots or wrappers; we are engineering the nervous system for the next generation of autonomous machines that live and act in the real world.

Sep 2025 – Feb 2026 · 5 mos · Paris, Île-de-France, France · Remote

• Friday was an autonomous AI agent designed to operate a computer exactly like a human user. Unlike chat-based assistants that only generate text, Friday could see the screen, understand the user interface, and actively control the mouse and keyboard to execute complex workflows across multiple applications.
• I engineered a vision-based control system that allowed the agent to perceive the macOS environment in real-time. The architecture featured an orchestration engine capable of breaking down high-level user intents into executable steps, delegating specific sub-tasks to specialized models like Perplexity for research or Gamma for content creation. This allowed Friday to automate end-to-end processes such as creating slide decks, managing emails, or organizing local files without human intervention.
• The project successfully demonstrated the potential of large action models (LAMs) and local-first AI. However, I concluded development in January 2026 to focus fully on the harder and more impactful problem of embodied intelligence at 20n Research Laboratory. Friday remains a testament to my ability to build complex, agentic software systems from the ground up.