AI Safety & Interpretability · Penn State
Junior at Penn State studying Applied Data Science. I work on interpretability and safety — trying to figure out what's actually happening inside language models.
I started college as a mechanical engineering major before switching to Applied Data Science at Penn State's College of IST. That pivot led me toward AI research — specifically interpretability and safety, trying to understand what's actually going on inside language models.
Recently co-first authored a paper on whether geometric properties of SAE decoder vectors can predict feature steerability — accepted to the ICML 2026 Mechanistic Interpretability Workshop. Now working on two new threads: applying interpretability methods to understand how models handle private information, and building practical mech interp tooling for finance. Also attending BlueDot's AI Safety program — thinking carefully about alignment and what it actually takes to make these systems safe.
Outside of research, I play poker with friends, play chess, listen to a lot of music, and just hang out. Originally from Nagpur, India. I also love astronomy and astrophotography — you can see some of my shots here.
Geometry Predicts SAE Feature Steerability
Algoverse AI Research · 2025 – 2026
Can geometric properties of SAE features predict how steerable they are — before you ever run a steering experiment? We show that decoder-space geometry (neighbor density, max cosine similarity) ranks features by steering cost, replicating across Gemma-2 scales, SAE widths, and architectures.
NLA Probing for Contextual Integrity
Python · PyTorch · June 2026 – Present
Exploring how language models internally represent contextual privacy norms using Natural Language Autoencoders — and whether models “know” when information sharing violates context-appropriate norms.
Applying mechanistic interpretability tools to finance applications.
Notes on what they do
and why they matter.
Click any entry to expand.