ML / Research Engineer
CEA Paris-Saclay · Apple
First-author work on Vision Transformer token pruning. Currently building planning and self-play agents from first principles.
Paris, France · Open to ML / Research Engineering opportunities
Token merging · keep ratio
98 / 196. Similar neighbors merge into larger blobs.
Selected work
Reinforcement Learning · Search · PyTorch
Rebuilding the path from classical search to AlphaZero — environment, agents, and evaluation in one harness.
Minimax. Searches future states with alpha-beta pruning.
Query
Does spatially coherent token selection change robustness under pruning?
↓ retrieved evidence
Passage 1
high relevance
Superpixel-based selection preserves neighboring tokens that independently sampled scores discard.
Passage 2
medium relevance
Accuracy at high keep-rates is close; the gap appears in dense prediction and under perturbation.
Passage 3
low relevance
A FLOPs reduction is reported, but latency is measured separately.
↓ generated answer
Yes, when the evaluation includes dense prediction and robustness, not only classification accuracy. [1][2]
RAG · LLM Systems · Evaluation
Question answering over papers, with answers grounded in retrieved passages and citations. Retrieval and generation are scored separately.
Hit@5
1.00
MRR
0.97
Development set
20
hand-labelled questions
Development set, not a production benchmark.
Experience
CEA Paris-Saclay
Research — Efficient Vision Transformers
Token pruning, token merging, mixed precision.
2025–2026
Apple
Camera Hardware Engineering
Image processing, optical measurement, lab automation. One system cut measurement runtime by about half.
2023–2024
Toulouse INP – ENSEEIHT
Engineering Degree + MSc AI
Ranked first in the final AI specialization year.
Graduated 2026
Research
VISAPP 2026 · First author
Studying whether spatially coherent token selection improves the robustness of Vision Transformer pruning across classification and dense prediction.
EEAI 2025 · Second author
A taxonomy of ViT pruning methods, with selected token and patch merging strategies measured on an NVIDIA Jetson Orin AGX. Fewer tokens do not always mean lower latency.
Token keep-set
Adjacent tokens kept together. Schematic, not a figure from the papers.
About
I trained as an electrical and electronics engineer, spent a gap year on two Camera Hardware internships at Apple, then shifted my final year toward machine learning at ENSEEIHT and joined CEA Paris-Saclay.
My current interests include representation learning, reinforcement learning, world models, and AI systems that have to operate under real constraints.
Currently
Looking for ML / Research Engineering roles where I can work close to the models, experiments, and systems — from the training loop through to evaluation.