Machine Learning Postdoctoral Researcher · Truveta · Bellevue, WA
I am a machine learning researcher trained in measurement science. On Truveta's Intelligence ML team I work across the model lifecycle on longitudinal electronic health records. I train clinical foundation models, including reinforcement learning that steers generated patient journeys toward real ones. I build the benchmarks and evaluations that test them, and the pipelines (MCP tools, agent systems, dashboards) that carry them into clinical workflows.
Measurement science is the unusual part of my background. A Ph.D. in measurement and statistics (University of Washington) and postdoctoral training in Harvard's CARES Lab (Causality, Applications, and Research in Education and Statistics) gave me tools that machine learning rarely borrows: validity theory, latent variable models, reliability analysis, and the design of randomized trials. I have applied them to very different data, from student essays and psychological surveys to electronic health records.
Across all of it, I ask whether a number means what we think it means, whether it is a benchmark score, an LLM grader, or a model-estimated patient-reported outcome. I ask it at every stage, not just at the end. Validity is not a final step; it is a design constraint.
Recent
Multi-agent temporal reasoning over longitudinal EHR for multi-cancer early detection · open access
Oral presentation · American Society of Clinical Oncology (ASCO) · Chicago, IL
International Society for Pharmacoeconomics and Outcomes Research (ISPOR) · Philadelphia, PA
TrajOnco since published; KCCQ work ongoing
Selected Work
A training-free multi-agent LLM framework reaching AUROCs of 0.64–0.80 across 15 cancer types. Published in npj Digital Medicine (2026); oral presentation at ASCO 2026.
Estimating a patient-reported heart failure outcome from routine EHR data, and checking that the estimate means what the questionnaire measures. First-author preprint, medRxiv.
A fine-tuned LLM reached 73% sensitivity at 91% specificity across 1,953 patients from 30 US health systems — a population current screening guidelines exclude on age alone. ASCO & AMIA 2025.
Whether LLM graders are psychometrically sound enough for high-stakes assessment. First-author EdWorkingPaper (2025) and a three-part blog series; presented at AEFP and JSM 2024.