I study the foundations of reliable machine learning, with the goal of understanding model behavior and leveraging these insights to improve reliability in real-world settings. I’m broadly interested in understanding how models learn, the ways they fail, and how we can build more robust, fair, and rational systems. Most recently, I have studied geographical failures in vision models, including studying the widening progress gap between imagenet-based benchmarks and global, crowdsourced data (ICLR 2024), and investigating the mechanisms of geographical bias (Spotlight, NeurIPS 2023).

My current research interests are largely motivated by my experiences deploying ML-based tools in high-stakes settings. I completed my undergradate at Duke University, where I worked with Mark Sendak as part of the Duke Institute for Healthcare Innovation (DIHI). At DIHI, I worked on building risk prediction models for severe pregnancy complications, which are now integrated and in silent trials. I also worked with DIHI to implement a data quality assurance framework which improved model performance by better integrating clinical feedback into dataset design. While at Duke, I also worked at the Duke Center for Global Women’s Health Technologies on a self-screening device for cervical cancer designed for low-resource global settings, which earned a Best Research award at NIH’s IEEE HIPOCT Conference in 2019.