My interests in AI are pretty broad, but they mostly come back to perception and representation: computer vision, representation learning, and JEPA-style world models. Reinforcement learning and vision-language models are never far behind. My PhD focuses on one small part of all this: teaching models to read documents, where the layout, tables and figures matter as much as the text itself. Away from research, I just like to build things. I'm comfortable in several languages (Python and C++ among them) and enjoy computing at every level, right down to low-level and embedded work. And I'm just as comfortable on the theory side, happy to take a problem all the way down to its mathematical foundations.
We use RL with a dense reward to train a CTC-based encoder-only model to do scene text recognition, we use the plackett-luce distribution and ctc-loss to derive this reward, learning an explicit, inspectable reading order over the image patches.
Uses mechanistic interpretability to guide pruning and distillation of the Donut VLM, yielding Donut-MINT, a lighter variant that keeps DocVQA performance at a fraction of the compute and memory.
Presents visualisation methods that expose how multiple modalies interact in vision-language models, making modality collapse easier to read and diagnose.
- 2026-06-30 Winning HacktheWorld(s): JEPA & World Models