ML Researcher — Generative Models · Computational Biology · Agentic AI
I build models that learn the structure of noise and agents that act in the real world.
Currently finishing my MSc thesis at IETR CentraleSupélec Université Paris-Saclay on discrete diffusion for DNA sequence reconstruction, and working as a Research Assistant at MBZUAI (Zhao Lab) on single-cell computational biology for autoimmune diseases.
📍 France · Open to Research & Industry — Available August 2026
| Project | What it does | Status |
|---|---|---|
| TraceDiff — MSc Thesis | Discrete diffusion model that learns structured noise (insertions, deletions, substitutions) in DNA sequences. Alignment-marginalizing encoder with attention pooling. State-of-the-art on Clustered Nanopore Reads. | 📝 Manuscript in prep (NeurIPS/ICML workshop) |
| T-Cell Atlas for MS — MBZUAI | Single-cell atlas integrating longitudinal scRNA-seq (EAE mouse + human MS data). Clustering, trajectory analysis, and differential expression with Scanpy. | 🔄 Ongoing |
|
World model bridging digital perception and biological computation — CNN encoder + DishBrain-inspired spiking neural network (LIF neurons, STDP plasticity). World Models Spiking Networks STDP
|
8 Transformer attention mechanisms (MHA, GQA, MQA, SWA, DiffAttn, MLA, MoH, NSA) implemented from scratch. Controlled pre-training on FineWeb-Edu with 42 unit tests. Transformers PyTorch From Scratch
|
|
Production pipeline generating synthetic DNA datasets calibrated to 7 synthesis platforms × 3 sequencing backends (Illumina, Nanopore, PacBio). 3-layer error injection. Noise Modeling Bioinformatics MIT License
|
LLM-driven multi-agent clinical platform — automated API orchestration across NCBI, PubMed, ClinVar. Real-time variant annotation and pathogenicity scoring. Multi-Agent Genomics FastAPI
|
|
Persistent multi-agent customs clearance system — 3 chained Gemini primitives: live supplier calls, persistent case memory, computer-use portal navigation. Built at RAISE Summit × Google DeepMind. Multi-Agent Gemini Computer Use
|
VLM + SLAM spatial mapping + cross-session knowledge persistence on Unitree Go2. Built at HACK 2026 — ETH Zurich Robotics Club (selected from 500+ applicants). VLMs SLAM Embodied AI
|
- 🥈 2nd Place — EPFL Quantum Hackathon (Quantum ML for option pricing, QRC for stock prediction)
- 🧬 RAISE Summit × Google DeepMind — Multi-agent customs clearance system
- 🏥 SOPHiA Genetics × ETH Zurich — AI for Health Hackathon
- 🤖 HACK 2026 — ETH Zurich Robotics Club — Embodied multimodal agent on Unitree Go2
| MSc Engineering — Polytech Nancy, Université de Lorraine | Intelligent & Autonomous Systems · Thesis at CentraleSupélec |
| Exchange — ÉTS Montréal | Computer Vision & Signal Processing |
| CPGE — Université de Lorraine | Mathematics & Computer Science |
Research Diffusion Models · Transformers · World Models · Agentic AI · Self-Supervised Learning
Representation Learning · Foundation Models · Spiking Networks · Clinical ML
Stack Python · PyTorch · CUDA · C/C++ · BF16/FP16 · Multi-GPU · Scanpy · FastAPI
Methods Ablation Studies · Controlled Experiments · Reproducible Research · Open Source
Building models that understand structure, not just data.