Ojasva Nema

Hi, I’m Ojasva.

I am an undergraduate at IIT Roorkee whose work revolves around Deep Generative Models, Mechanistic Interpretability, and Large Language Models. Currently, my work is focused on the interpretability side as a member of the Data Science Group (DSG). I lead a small research squad there investigating LLM internals, and we recently published work on how models route domain knowledge (check out our website).

Simultaneously, I work at the P-square Lab with Prof. Parikshit Pareek. We are investigating Bilinear MLPs to find more interpretable architectures and building a foundation model for amortized kernel hyperparameter discovery. We also recently finalized a Sparse Diffusion framework for high-energy physics (CERN).

Previously, I worked on Image Alchemy, a personalization method for diffusion models that I published at ICLR DeLTa workshop 2025. I also gained production-grade AI experience at Elimentary, where I engineered local LLM APIs and evaluation pipelines.

Off the clock, I’m usually reading about finance (I’m always open to book recommendations!) or watching anime. Always happy to chat about circuits, LLMs, or what’s next in AI!

(And seriously, what would you do if you had infinite compute?)