I'm a final-year computer science undergraduate at MAIT in Delhi, graduating in 2027. I work on AI safety, mostly on one question: when a method claims to read what's inside a model, how would we know if it's wrong? I build models where the answer is known in advance (a habit installed, hidden, or never there) and check whether the instruments find it. Usually they don't, and some of my favourite results are false positives we caught in our own work.
I co-founded Redarc Labs, a nonprofit that studies what deployed AI systems can do that their models alone can't. I'm also a research fellow at MARS (Cambridge AI Safety Hub), working on model organisms with hidden objectives, and a special projects fellow at Secure AI Futures Lab, which works on trustworthy AI for India and South Asia. This summer I was a Google Summer of Code fellow with HumanAI, writing pathfinding algorithms to work out how people moved through the El Bagawat necropolis in Egypt.
Papers
* equal contribution. NeurIPS workshop papers are non-archival.
- Known by Construction: Calibrating Model-State Instruments Against Models Whose State Was Installed, Hidden, or Never Present New in ML, NeurIPS 2026
- Dead Salmons Have a False-Positive Rate: A Calibrated Null-Model Test for Interpretability Interpretability as a Science, NeurIPS 2026
- Steered Affect Moves a Partner Agent's Activations, and the Logistic Probe That Reads It Returns a Tie-Break, Not an Estimate Interpreting Agent Behavior, NeurIPS 2026
- Toxin Circuits in ESM-2: Mechanistic Interpretability for Biosecurity Screening GenBio Workshop, ICML 2026
- Distance to the Boundary Is Not Steering Resistance Under review, UniReps
- Doubt Is a Decision: A Sparse Circuit Commits Language Models to "I Don't Know" arXiv preprint, coming soon
Before this
Technical AI Safety fellow at the AI Safety Initiative, Georgia Tech (2026). BlueDot Impact courses in technical AI safety and biosecurity (2026). 6th of 1800+ in the Haas F1 Challenge, on video anomaly detection (2025).