My research is driven by low-dimensional principles of intelligence, seeking the foundational blueprints that enable scalable, generalizable AI. This perspective (see a "philosophical" piece I wrote for the CPAL conference) has guided my work across many domains, aiming to move beyond brute-force scaling toward more structured and principled learning systems.

 

At the university level, we have recently focused on how training dynamics can inherently discover low-dimensional inductive biases, such as sparsity, low-rank manifolds, and algebraic symmetries, shaping more efficient paradigms for pre-training, supervised fine-tuning, and reinforcement learning. We further study how these structures manifest at test time as self-organized, compositional mechanisms that support inference, reasoning, and agentic planning. Finally, we are interested in how such structural invariants translate into real-world deployment, enabling resource-aware and trustworthy AI in high-stakes domains.

 

Historically, my work has included deep learning theory, learning-augmented optimization, sparse coding & inverse problems, and visual restoration & understanding — areas in which I remain (somehow) engaged.

​

In parallel, my industry experiences have opened several new, distinct research directions inspired by practical deployments and large-scale systems: geometry deep learning and graphs (with Amazon),  video generation (with Picsart), and foundation model training for trading (with XTX Markets)