About me
I am an Applied Scientist at AWS, where I work on research problems at the intersection of machine learning and health. I am extremely passionate about healthcare and the potential that technology has in revolutionizing it—especially in the area of behavioral health, where much of the critical data remains uncollected. With the increasing presence of devices and sensors in our daily lives, I believe we now have the tools to bridge this gap and design better, more responsive systems for care.
Prior to this, I completed my Ph.D. at the Georgia Institute of Technology, where I was advised by Prof. Jim Rehg. My research focused on machine learning methods for mobile health, including the use of physiological signals, digital biomarkers, and behavioral modeling.
Outside of work, I enjoy being outdoors—hiking, biking, and running. I also enjoy theatre performances and am proficient in Indian classical music.
Projects
Semantic Highlighting for Query-Relevant Text
When users need to find answers quickly in long documents such as physicians searching through medical records, highlighting the most relevant text is key. We fine-tuned a language model with a lightweight prediction layer to identify query-relevant spans, enabling faster and more focused information retrieval. The model generalizes well across domains and is now part of OpenSearch 3.0. This approach is especially valuable in clinical settings, where timely access to specific information can impact decision-making.
Colloquial Radiology Reports with Large Language Models
To improve patient comprehension, we developed a method to automatically translate radiology reports traditionally written in clinical jargon into more accessible, patient-friendly language. Our approach leverages large language models and was evaluated on accuracy, readability, likability, and potential for harm. This project was conducted in collaboration with UCI Radiology and is detailed in our JAMIA publication.
Domain Adaptation for Pulsative Physiological Signals
Wearables and mobile devices offer continuous insights into human physiology, but predictive models often struggle with domain shifts - differences in sensor type, placement (e.g., ECG leads, PPG sites), or population between training and deployment. To tackle this, I designed self-supervised learning tasks that capture the underlying structure of physiological signals. This approach significantly improved model transferability, enabling more robust and accurate health monitoring across varied real-world conditions.