I am a first-year PhD student in Computer Science at Harvard University, advised by Prof. Flavio du Pin Calmon in the Information Theory Lab.
Before coming to Harvard, I completed a BS in Applied Mathematics and an MS in Computer Science at UC San Diego. During my time at UCSD, I was fortunate to be advised by Prof. Berk Ustun, and to work with Prof. Benjamin Smarr, Prof. Zhiting Hu, and Intel Corporation.
Outside of research, I like to run, and watch sports (mainly baseball and basketball).
Some topics that I am currently interested in:
Models are typically tested on data gathered in advance, which says little about how they behave once the world around them moves: when people react to their decisions, or when they are used somewhere different from where they were built. I want to build evaluations aimed at specific questions rather than general benchmarks, and make them usable by the people who need them most, including the individuals affected by a decision, who rarely have the access required to check it.
Most systems are designed to run on their own, but safety in high-stakes settings depends on people staying involved, correcting what the model got wrong or taking over when it should not be deciding at all. I want to study when a model ought to hand a decision back, based on how much harm a wrong answer would do rather than only on how uncertain the model is, since uncertainty-based rules tend to fall hardest on minority groups.
NeurIPS Evaluations & Datasets 2026
Concept bottleneck models are hard to evaluate because almost no real dataset carries concept labels, so we generate synthetic benchmarks that do, with explicit control over the properties that make these models succeed or fail.
arXiv preprint, 2026
Deployed models face inputs that shift through noise, strategic gaming, or ordinary interaction, so we estimate how often and how much predictions change under those shifts, with statistical guarantees that support real validation.
ICLR 2025
Standard explanations often tell people to act on features that would never flip their prediction, so we score each feature by the probability that acting on it actually changes the outcome.