Research experience

• Developing modular AI-agent harnesses for scientific discovery, comparing orchestration, structured search, memory, and role-based components under matched backbone, interface, and compute budgets.
• Studying where humans should intervene in agentic scientific workflows, with focus on failure modes, task boundaries beyond current agent capability, and help-seeking policies that elicit useful human input.

Understanding LLM dynamics as a function of pretraining data

• Studying how pretraining data shapes clinical LLM behavior and probabilistic knowledge.
• Developed corpus-based probabilistic baselines that complement LLM diagnostic reasoning.

Interpreting Dataset Shift in Clinical Notes

• Developed SIReNs to explain dataset shifts in clinical notes; extending this work to agent-based shift analysis.

Large Language Models for Time Series Forecasting

• Developed GPT4MTS to combine time series and text for multimodal forecasting.
• Developed TEMPO with trend decomposition and adaptive prompting for forecasting under distribution shift.

Initializing Adapters with Improvised Knowledge for Multimodal Continual Learning

• Studied adapter-based knowledge transfer for continual visual question answering to improve forward transfer and reduce forgetting.