martin.legrand

Arxiv: https://arxiv.org/abs/2603.28986

A multi-agent AI framework for scientific research. Mimosa-AI is designed to autonomously generate hypotheses, design experiments, and analyze results, accelerating the pace of discovery in various scientific domains. It use innovattive, Neuroevolution inspired techniques to evolve its own architecture and learning strategies over time, Build in my current position at CNRS HolobiomicsLab .

23k+ stars · Python · LLM Agents

A fully local alternative to Manus AI. Autonomous agentic system capable of complex task execution without cloud dependencies. Designed for privacy-conscious users and organizations requiring on-premise AI solutions.

In progress

A extremely frugal and efficient LLM agentic framework designed to run fast on low-resource devices. The idea: the harness walks a tree of decisions and renders each one as a numbered menu; the model answers with ~3 tokens. Everything deterministic (parsing, page rendering, notes, undo) is code; the model is only the policy oracle.

C++ · OpenCV · Computer Vision

Real-time simultaneous localization and mapping using a single camera. 3D reconstruction pipeline with feature detection, pose estimation, and sparse point cloud generation for applications such as robotics.

PyTorch · Generative AI · Audio

Generative adversarial network for synthetic audio generation. End-to-end pipeline for training and inference with MLOps integration for reproducible experiments and model versioning.

PyTorch · RL · OpenAI Gym

Collection of reinforcement learning solutions for classic control problems: Cart-Pole, Lunar Lander, Bipedal Walker. Implements DQN, PPO, and policy gradient methods with detailed performance analysis.