Experience

Feb. 2026 - Present
Working at Project Numina, a non-profit advancing open-source AI for science. I am developing an optimization system that combines Reinforcement Learning with Graph Neural Networks to tackle mathematically complex problems involving symbolic computation and poly-exponential complexity growth. The training pipeline scales RL agents across high-dimensional search spaces using Ray for distributed multi-GPU training. (Temporary NDA — further details available upon request.)

5 months
Sep. 2025 - Jan. 2026
Modeling the acquisition of language capabilities in Large Language Models. I implemented a time-series based surprisal analysis on a controlled dataset of grammatical vs. ungrammatical constructions to investigate the dynamics of grammar acquisition during LLM training.

I've done a five-month internship at the Institute of Neuroinformatics in Zurich as an ETH Zurich student. My research focused on artificial neurogenesis in the context of hardware-constrained continual learning. Specifically, my team previously developed a neuromorphic hardware chip called MOSAIC for in-memory computing. I have been investigating the development of a neurogenesis approach to continual learning, designed to be implemented on MOSAIC chips.

The goal of this research internship was to delve deeper into research project below.

During a 3 months intenrship at CERN, I conducted an in-depth investigation of Domain Adversarial Neural Networks, including a comparative review of various architectures and the refinement of the most relevant one to deal with highly imbalanced High Energy Physics datasets. More precisely, I embedded the domain information of each sample into the input space, which enabled the model greater flexibility in building the decision boundary.

The Fair Universe project primarily aims at mitigating systematic uncertainty in High Energy Physics through the elaboration of machine learning challenges.

I have been involved in this project part-time for a preliminary exploration of baselines to tackle systematic uncertainty. Additionnaly, I have been leading a team of 4 in the conception of a machine learning challenge which has been solved by students from a class of 20.