Currently seeking full-time positions

Pronoy Das

Aspiring AI Safety Researcher · Computational Physicist · ECE PhD Candidate, Purdue University

I am a computational and theoretical physics PhD candidate at Purdue, switching my career direction towards AI safety. I am currently working on the interpretability of models during the RL training process. Previously, I worked on physics problems realted to quantum optics + field theory. I have also worked on optical computing hardware at Sandia National Laboratories, and GPU (CUDA) based high-performance computing pipelines for scientific simulations and imaging.

What I do

AI Safety

Model internals during RLVR

As a part of the BlueDot Impact's AI project, I studied misalignment during RLVR training over intermediate checkpoints. I discovered that RLVR broadly introduceds self-doubt in the model's behavior, rather than behavioral misalignment (read the article in the 'Articles' section).

Computational Physics

Computing with physical systems

With Zetascope, I developed dedicated algorithms and corresponding CUDA C++ kernels for near real-time faint-object detection pipelines.

Quantum

Quantum theory

I developed the quantum theory of light-matter interaction in chiral crystals, and of the angular momentum of structured light, leading to several journal articles, including a Phys. Rev. B Letter (Editor's Suggestion). I also co-authored Purdue-PicoMax, a C++ electromagnetic simulation package built for the theory.

Selected publications

All publications →  ·  Google Scholar ↗

Awards

Experience & education →

Contact

The best way to reach me is by email. I read both inboxes daily, and I reply quickly to recruiting and collaboration inquiries.