Academic Projects

Reinforcement Learning

• Robot Control:
  • Goal: Training a wheeled-bipedal robot to recover from pushes using PPO, reward shaping, curriculum, inductive bias.
  • Supervised by: INRIA Willow team (S. Caron)
• Traffic Optimization using Autonomous Vehicles:
  • Goal: Exploring how DQNs (and PPO) can learn to control autonomous vehicles to avoid traffic jams
  • Inspired by: CIRCLES project conducted by Berkely, Rutgers, ENPC…
  • With: 3 mates from ENPC. Supervised by 2 professors
  • Resources: Report ; RL Environment

Deep learning, physics, vision

• Physics Informed Neural Networks (PINNs):
  • Goal: Solve Helmholtz PDEs with PINNs
  • With: 2 mates from ENPC and 2 scientists from Airbus
  • Resources: Poster ; Report ; Code
• Neural Radiance Fields (NERFs) for Scene Video Synthesis:
  • With: Two mates from ENPC. Supervised by Prof. Vincent Lepetit
  • Resources: Slides; Code
• Self-supervised learning for point clouds:
  • Goal: Paper study
  • Context: MVA geometric deep learning course

Probabilistic methods

• Denoising Score Matching:
  • Goal: Reduce sampling bias
  • Context: MVA probabilistic graphical models course

NLP

• RNN-based Neural Machine Translation Model:
  • Goal: Trained an LSTM-based machine translation toy model (20M parameters) and integrated it in a web interface
  • With: Individual project at ENPC
  • Resources: Code

Optimization & Operations research

• Offshore Wind Electrical Network Optimization:
  • Goal: Optimizing an offshore wind electrical network for RTE, FR
  • Competition: KIRO 2023 (Project), with 2 mates from ENPC
  • Using: Python, Julia, C++, Gurobi (MIP), Hexaly, S.Annealing
  • Resources: Code
• Manufacturing Chain Optimization:
  • Goal: Optimizing a car manufacturing chain for Renault, FR
  • Competition: KIRO 2024 (Hackathon), with 2 mates from ENPC
  • Using: Python, C++, Julia, Gurobi (MIP), Hexaly, S.Annealing
  • Resources: Code

Internship Projects:

• MLO Lab, CHE 🇨🇭
  • See details in experience, here.
• TII, UAE 🇦🇪
  • Knowledge Distillation and model pruning strategies.
  • Large-scale optimization algorithms and scaling laws for parametrization of LLMs.
  • Example contribution to an open-source Nvidia framework: PR
  • Falcon3 family is finally released (SOTA)!
  • More releases coming in 2025…

P.S

Some code links might not be accessible.
This means generally that I (we) am (are) working on cleaning them.
I may also have changed the repo name (sorry). But I’ll try to update every once in a while.