I am an applied mathematics engineer from École polytechnique, based in Paris, now specializing in AI at the MVA (ENS Paris-Saclay), with a focus on optimal transport, deep learning, and image processing.
I also enjoy teaching and am open to new opportunities—feel free to reach out!
Last year, at the Centro de Modelamiento Matemático (CMM), I worked on online algorithms for combinatorial auctions with J. Correa and A. Cristi. Before that, at Inria, I built tropical support vector machines with X. Allamigeon, S. Gaubert, and T. Molfessis.
In 2023, at Exotec, I doubled the speed of thousands of warehouse robots by developing a new autopilot. Before that, at the Centre de Mathématiques Appliquées (CMAP), I simulated population dynamics with V. Bansaye, M. Breden, M. Grau, and D. Sbeiti.
I contribute to the Python Optimal Transport (POT) library.
In online combinatorial auctions, m objects are assignable to n agents who arrive sequentially, in an adversarial order. Each agent has a valuation for each possible bundle of objects. The aim is to distribute these objects on the fly to maximize global welfare. Are there prices that guarantee 2/3 of the offline optimum? I obtained partial results on this challenging open problem for a small number of items and in the simplified context of max-min greedy matching.
We developed max-margin SVMs with efficient training and inference, based on tropical geometry and mean payoff games, with theoretical guarantees for margin optimality, all-vs-all multi-class classification, and the ability to perform general piecewise linear classification. [repo]
I designed, tested, and deployed an efficient model predictive controller for the Skypod robot, enabling it to intelligently anticipate its trajectories and learn about its environment's imperfections. Using its physical model, the robot is now able to calculate where a series of commands will take it, and choose the optimal trajectory in the long run. Thanks to my controller, we were able to safely double the robots' maximum speed to 4 m/s, enabling them to carry out far more orders for hundreds of customers in industry, healthcare, retail...
We studied the dynamics of two rival species and observed that the cross-diffusion term in the SKT model induced spatial segregation at equilibrium. To this end, we developed a fast solver of the underlying non-linear differential equations. We then showed that this continuous model was the limit of a Markovian random process, which we also simulated. [repo] [pypi]