Fourth-year PhD candidate in the Theory of Machine Learning Lab at EPFL, advised by Nicolas Flammarion.

I work on understanding learning systems from first principles: training dynamics, generalization, and the structure of sequential data, with a focus on transformers and autoregressive processes.

Website · Publications · Scholar · CV

Most of my experimental work lives in a few long-running project repositories:

• hmm-attention — hidden Markov models with hierarchical structure, attention probes, and dissection of internal computations.
• ar-classification — autoregressive models and transformers on structured sequence classification tasks.
• ar-transformers — autoregressive reasoning tasks on transformers.
• ar-linearsubspace — autoregressive linear models: stability, estimation, and sample-complexity experiments. Supports Long-Context Linear System Identification, ICLR 2025.

• sparse-attention-dynamics — Incremental Learning of Sparse Attention Patterns in Transformers, ICML 2026.
• latentclr — LatentCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions, ICCV 2021.
• phuber — [Re] Can gradient clipping mitigate label noise?, ReScience C 2021.

• lean4ml — Lean 4 formalization of optimization results. Contributor.
• Winter School on Synchronization 2020 — three teaching notebooks covering numerical integration, chaotic dynamics and the Lorenz system, and synchronization on networks. Refactored from the school's hands-on material into composable primitives: pluggable integrators and network dynamics assembled from subsystem, coupling, and topology. Co-organizer of the school; also built its website.

Before and during my bachelor's, I was an indie game developer and released games on the iOS App Store (RitmiSay, Tipatip), and built backend, iOS, and ML systems, first as a freelancer and later as co-founder at RateMe, a computer-vision startup. The code for these projects is mostly proprietary and not listed. Here are some public projects:

• Actopus — full-stack TypeScript activity platform, nine-person team. Top contributor, 255 commits.
• modbus — a lightweight Modbus slave library for Arduino, built around an abstracted communication layer so serial and ethernet transports subclass a common core, with configurable register memory and user-definable function codes.