About me

I am an Applied Scientist at AWS Health AI, where I work on responsible AI for multiple healthcare agents.

I recently completed my PhD at Cornell University, where I was advised by Mert Sabuncu. My doctoral research focused on developing time-series and computer vision methods for predicting clinical outcomes from longitudinal multimodal biomedical data.

During a Summer 2025 internship at Amazon AGI, I developed a novel REINFORCE-based RLVR algorithm for reasoning LLMs that achieves state-of-the-art performance across multiple benchmarks.

During a Summer 2024 internship at Microsoft AI, I worked on responsible AI research, developing supervised fine-tuning and alignment methods to improve LLM safety and usefulness.

For more details, please refer to my resume.

Highlighted Publications

LoMaR: Longitudinal Mammogram Risk Prediction

MICCAI 2024
Extends a state-of-the-art breast cancer risk model to ingest an arbitrary number of longitudinal mammograms. Longer histories of prior annual mammograms significantly boost the accuracy of predicting future breast cancer risk.

Code

DISPO: Decoupled Importance Sampling-weighted Policy Optimization

AISTATS 2026
A REINFORCE-style RL algorithm for LLM mathematical reasoning that clips importance-sampling weights with separate, independently tunable bounds across four policy-update regimes, matching PPO-style stability while training as efficiently as REINFORCE.

Code

A Deep Survival Model for Predicting Alzheimer's Diagnosis Based on Multi-Modal Longitudinal Data

Big Data Mining and Analytics, 2026
A Transformer-based survival model (LSM-AD) that integrates cognitive assessments and MRI biomarkers across a patient’s visit history to predict future Alzheimer’s diagnosis, improving AUROC by up to 15% over prior state-of-the-art.
Code

Paper

POROver: Improving Safety and Reducing Overrefusal with Overgeneration and Preference Optimization

ICML 2025
Shows that overgenerating instruction-finetuning data with stronger teacher models improves the safety/usefulness trade-off, and introduces POROver, a DPO-based alignment strategy that reduces LLM overrefusal while preserving safety.

Code