Research overview

Auditable AI
for Biomedicine

Reliable evidence. Controllable reasoning. Auditable action.

I develop AI systems that can reason, adapt, and act in complex real-world environments—turning heterogeneous evidence into trustworthy decisions, scientific discoveries, and better paths to treatment.

View publications Contact me

Heterogeneous biomedical evidence

Imaging

EHR

Notes

Genomics

Knowledge

Reliable learning

Imperfect evidence

Controllable reasoning

Steer and adapt

Agentic systems

Act and verify

Auditable AI for real-world biomedicine

Better clinical decisions

Scientific discoveries

More effective treatments

Reliable Learning from Real World Data

Real world data are noisy, incomplete, and heterogeneous. I develop methods that train model reliably when training data are imperfect.

RobustnessGeneralizationReal-world data

Selected work

Controllable and Auditable Reasoning

Powerful models are not enough. We need to control how they reason, adapt their behavior at inference time, and understand why their outputs change (with human language as feedback).

SteeringAdaptationInterpretability

Selected work

Agentic AI with Verifiable Actions

AI systems that gather evidence, use tools, and take multi-step actions while preserving a transparent and inspectable decision trail.

Tool useMulti-agent systemsSelf-audit

Selected work

OctoTools: an agentic framework with extensible tools (NAACL, Best Paper)
CEREBRA: a multidisciplinary agentic AI system for dementia characterization and risk assessment (Under Review @ Nature Medicine)
In-the-flow agentic system optimization for effective planning and tool use (ICLR, Oral)

From research to impact

Turning AI into real biomedical value

We apply these ideas to high-stakes problems in biomedicine, where reliability and auditability are requirements, not preferences.

Clinical decisions

Disease risk stratification

Multimodal reasoning across imaging, EHR, clinical notes, and laboratory evidence.

Selected work

CEREBRA: an agentic AI system for dementia characterization and risk assessment (Under Review @ Nature Medicine)
Generalizable deep learning model for early Alzheimer's disease detection from structural MRIs (Scientific Reports)
Time-Series Analysis via Low-Rank Matrix Factorization Applied to Infant-Sleep Data (NeurIPS ML4H)

Treatment planning

AI-guided radiotherapy

Turning clinical intent into auditable, deliverable treatment plans.

Selected work

GPT-RadPlan: automated radiotherapy planning with large language models (Physics in Medicine & Biology)
Automated Treatment Planning Guided By a Context-Aware Foundation Model (AAPM Best in Physics Therapy)
A plugin for automated treatment planning in Eclipse TPS based on large language models (Medical Physics)
TextGrad: Automatic "Differentiation" via Text (Nature)

Scientific discovery

AI for Biomedical Discovery

Turning heterogeneous evidence into testable hypotheses.

Selected work

We are interested in AI systems that connect evidence across clinical data, imaging, literature, and biological measurements to generate and refine scientific hypotheses. (Ongoing)