Research background
How I approach evidence
Experimental design, statistical and causal inference for observational data, and scientific computing for large-scale health data.
Research
Evaluating models as experiments
A benchmark is an experiment. The model is a population, the prompt is a treatment you assign, and the score is an estimator with sampling error and measurement error. So I ask what epidemiology asks of any study: is the sample representative, is the comparison powered and fair, is the reference standard reliable, and could selection or leakage explain the result?
// meta-analysis published in npj Digital MedicinePIHF-MCP: agents that learn from expert investigation
Policy Iteration with Human Feedback turns the way experts investigate model failures into a reusable, recorded policy. An agent executes that policy while the model weights stay fixed. PIHF-MCP automates the loop; in the preprint, policies developed from a few labelled examples improved performance on unseen cases and across models.
// preprint on arXivliteOdyssey: rare-disease diagnosis you can inspect
liteOdyssey consolidates model failures and expert corrections into a clinician-gated natural-language policy for rare-disease diagnosis. The system is evaluated on held-out cases across benchmarks and model families and, after adaptation, in the Undiagnosed Diseases Network with blinded physician adjudication. Reviewers can run it in a live workbench: question the system, inspect its tool trace and reasoning, and test where the policy holds or fails.
// liteodyssey.org // demo video // live workbench (request access) // preprint on arXivBelow the Surface: hidden bias in large-scale health data
My dissertation models the biases that sit below an estimate: how outcomes and exposures are measured, who gets selected into a cohort, and how far an unmeasured confounder could move the answer. The setting is NASA-supported research on low-dose radiation and Parkinson's disease in Navy submariners. The same habits run through my medication-safety work, from nationwide new-user cohorts to depression risk with 5-alpha reductase inhibitors.