Hi! I am a final-year PhD candidate in Computer Science at Stanford University. I am co-advised by Dan Boneh and Dorsa Sadigh. I am also an AI Science Writer at Anthropic.

My research sits at the intersection of machine learning, safety, and security, with a focus on reducing both the computational and cognitive costs of verifying modern AI systems.
I take an end-to-end approach to improve our ability to audit, understand, and learn from model behavior across diverse domains. Some recent projects I've worked on include:

1. Optimistic verifiable training and Hawkeye, methods for auditing large neural network computations that can vary across different GPUs (NeurIPS '24, MLSys '26).
2. One of the first controlled user studies of security vulnerabilities in AI-assisted programming (ACM CCS '23, TechCrunch).
3. Z-COACH, a shared-autonomy framework for teaching humans motor skills from expert AI policies, developed with Toyota (HRI '25).
4. Policy Learning with a Language Bottleneck, a method for learning interpretable RL policies that improve human-AI coordination and generalization (TMLR '26).
5. Pyrite, a benchmark and RL-trained agent that generates interpretable code to evaluate diffusion-predicted protein–ligand poses, outperforming human experts. (preprint).

My research has been generously supported by both the NSF GRFP and IBM PhD Fellowship.
During my PhD, I have interned at Genesis Molecular AI, visited the LINGO lab at MIT, and collaborated closely with Stanford's Computation & Cognition Lab.
I completed my undergrad in Computer Science and Creative Writing at Stanford, where I was a member of the Stanford NLP group and received the Ben Wegbreit Prize for Best Undergraduate Honors Thesis in Computer Science.

Updates