Projects

This page presents a selection of industrial and research projects I have worked on. Most were carried out at CEA-List and with its industrial partners, while others were conducted independently. They primarily focus on AI for numerical simulation and scientific computing, and also extend to other areas such as geolocation and natural language processing.

A scene simulated by LaGSplat. Dragging inside the video applies a force to the object, and every frame is generated in real time under that force.

Computer Vision · CEA-List · 2026

LaGSplat: Physics-Governed Interactive Simulation from Monocular Video

World Models
Gaussian Splatting
PyTorch

LaGSplat infers the physics of an object from a few seconds of monocular video. A visual encoder projects each frame into a low-dimensional latent space in which a Lagrangian Neural Network learns the equations of motion, and a Gaussian Splatting decoder renders the scene in real time. LaGSplat can extrapolate in time, replay the video from different initial conditions or even apply an interactive force to the scene, even though no force data was ever seen during training. Patent filed, as sole inventor. Project carried out as part of my work at CEA-List.

Interactive demo → Related publication →

Impostral landing page, the game that won us 3rd place at the Mistral Vibe Hackathon 2026.

Mistral Vibe Hackathon · 3rd place · July 2026

Impostral

Agentic AI
FastAPI
STT / TTS

Built Impostral at the Mistral Hackathon with my team: a multiplayer Turing test. A social deduction game where Mistral agents infiltrate a group of humans and try to blend in. Can you unmask the AIs before they convince everyone that you’re the AI?
🏆 1st of our task and 🥉 3rd overall. Awarded 1000 € in Mistral credits.

Play Impostral → GitHub →

Grasping operation by a soft gripper simulated in real time using a LEBNN.

European Research Programme · CEA-List · 2024–2026

ARISE

NVIDIA Omniverse
PyTorch

ARISE is a European research programme aimed at designing soft industrial robots operating in close contact with human operators. Robots are trained via reinforcement learning in a physically realistic virtual reality simulation developed by CEA. My LEBNN neural network architecture (published in JMPS 2025) is used to simulate hyperelastic components of the 3D scene in real time. Project carried out as part of my work at CEA-List.

Project website → Related publication →

Contact simulation between a soft gripper (simulated by a neural network) and a rigid object (simulated by XDE).

Numerical Simulation · CEA-List · 2025–2026

Neural Network / XDE Coupling

C++
libtorch
PyTorch

Development of a real-time coupling method between a neural network surrogate and the XDE physics engine, implemented in C++ with libtorch. The neural network handles computationally intensive components, allowing the full simulation to run in real time. Project carried out as part of my work at CEA-List.

Interactive demo →

Generic illustration generated by Mistral AI - NDA compliance.

Data Science · Freelance · 2025

AI for indoor geolocation - Wheere

PyTorch
ONNX
Signal Processing

Freelance data science mission dedicated to indoor geolocation under severe noise and interference. Designed and implemented a full pipeline from scratch to adress this ambitious problem, including data generation, signal preprocessing, neural network architecture and post-processing. Developed an highly specialized neural network architectures operating jointly in the time and frequency domains, combining signal processing, spectral analysis, and tailored mathematical modeling. The resulting approach reduced localization error by 70% in the most degraded scenarios.

Wheere website →

Generic illustration generated by Mistral AI - NDA compliance.

Industrial Consulting · CEA-List · 2025

AI surrogate model for design optimisation - CEA-List × Forvia

PyTorch
ONNX

Neural surrogate trained to predict physical fields on an automotive component, under severe data scarcity constraints. The model is integrated into a design optimization loop, replacing computationally expensive finite-element simulations. Project carried out as part of my work at CEA-List.

3D reconstruction of myself climbing, generated from a single photo. V7/7a+ btw.

3D Vision · Freelance · 2024

POC Arkose

Python
3DGS

Proof-of-concept for Arkose, a climbing gym chain: reconstructing a climbing wall in 3D from a single photograph, with no depth sensor or LiDAR. Monocular depth estimation feeds a 3D Gaussian Splatting pipeline that turns one image into a navigable scene.

Arkose website →

Generic illustration generated by Mistral AI - NDA compliance.

Industrial Consulting · CEA-List · 2024

Crash Test Simulation from Scarce Data - CEA-List × Renault

GNN
PyTorch

Development of a neural network model for crash test simulation under severe data scarcity constraints. Designed a parsimonious architecture tailored to the low-data regime, replacing computationally expensive finite-element simulations. Reduction in computing time by several orders of magnitude. Project carried out as part of my work at CEA-List.

Generic illustration generated by Mistral AI - NDA compliance.

NLP · Freelance · 2023

NLP for Intellectual Property

LLM
Word2Vec
NLP

Freelance NLP mission. Developed a patent drafting assistance tool combining Word2Vec embeddings and LLM-based generation to suggest claim formulations from a technical specification. The pipeline extracts key concepts from a technical document and outputs a structured claim tree.

Example of causality graphs and the associated DGLSTM network employed to extract information from ST code.

Deep Learning · SENS · Crestic · Université Champagne-Ardenne · 2022

PLC Language Processing

LSTM
PyTorch
NLP

Research mission carried out through the ENS junior enterprise. Design of a neural network architecture tailored to PLC ladder logic language processing, combining graph-based and recurrent components. Led to a peer-reviewed publication (MDPI Processes, 2023).

Related publication →

Predicted and simulated fluid velocity fields, along with the error field, for the von Kármán test case using our GNN architecture.

Research Internship · EDF R&D · 2021

Graph Neural Network for Fluid Dynamics - EDF R&D

PyTorch
GNN
SplineConv

M2 research internship at EDF R&D, focused on model-order reduction applied to fluid dynamics. Designed a graph neural network architecture combining SplineConv layers and a hierarchical structure to predict fluid flow fields at a fraction of the cost of classical solvers. Presented at the NeurIPS 2021 Workshop on Machine Learning and the Physical Sciences.

Related publication →

Example result from the coupled simulation. The joint learned by the neural network is present 256 times in each grid of the fuel assembly.

Numerical Simulation · CEA · Centre Borelli · 2020

Neural Network / Cast3M Coupling for Nuclear Assembly Simulation

LSTM
Cast3M
PyTorch

Research project carried out as part of the ENS AI Research Year. A LSTM network was trained to learn the behaviour of a mechanical joint present in very large numbers in nuclear fuel assemblies. The model was integrated into a Cast3M finite-element simulation via a PyTorch coupling, reducing computation time by 95%.