Computational biology × AI safety and evaluation
I build benchmarks and reproducible systems that test whether models ground specialist outputs, allocate verification where it matters, generalize across real biological data, and stay calibrated when those outputs reach a real laboratory.
Senior Research Associate and Director for Spaceflight · Mason Lab, Weill Cornell Medicine
Selected work
Seven public projects spanning model grounding, verification-aware agents, mission-shift generalization, drug-discovery decisions, laboratory execution, evidence across agent handoffs, and safeguard calibration. Each links claims to inspectable code, data, or result artifacts.
A measurement-first map of whether language models ground the content of biological representations—sequence, structure, identifiers, and numeric predictions—or rely on labels and familiarity. It connects measured gaps to a practical train, retrieve, or orchestrate decision map.
A verifiable-reward benchmark for a central agentic decision: when a biology foundation model may be wrong, should the agent trust it or spend budget on a real differential-expression assay?
Mission-held-out transcriptomics benchmarks for a hard biological generalization question: can a model trained on one spaceflight study recognize a signal in a mission it has never seen?
An auditable clinical and regulatory decision benchmark with callable data adapters, verifier contracts, an installable scorer, and reproducible clinical-trial decision labels.
Current scope: a retrospective benchmark plus one audited end-to-end disease slice—not a finished eight-stage discovery platform.
An Inspect AI environment that evaluates agents executing benign molecular-microbiology protocols inside a seeded stochastic simulator, with deterministic four-axis trajectory scoring.
A prespecified test of whether scientific evidence survives handoffs between agents, or whether the last agent in a chain ends up citing its predecessor instead of the original source, and how far a structured claim record and a runtime gate recover what is lost.
A controlled study of whether safeguards stay calibrated when the same science changes form, moving across names, database identifiers, sequences and annotations, and when delivery moves from a chat window to a tool call. Released with an outcomes-only public dataset.
Research program
My evaluation work is grounded in computational biology: human spaceflight multi-omics, single-cell and spatial genomics, perturb-seq, and the practical constraints of scientific workflows.
Featured benchmark
A multi-omics AI benchmark built from public spaceflight biomedical releases across Inspiration4, the NASA Twins Study, and JAXA Cell-Free Epigenome data.
How I work
Research foundation
I serve as Senior Research Associate and Director for Spaceflight in the Mason Lab at Weill Cornell Medicine, where I connect biomedical discovery with rigorous AI evaluation.
Individual scholarship supporting a transition into AI-bio safety research.
Compute support for research on evidence-aware biological language models.
Principal investigator allocation for scalable computational research.
Orbital launch funding for spaceflight biomedical research.
Connect
Open to research conversations across AI for science, evaluation, post-training, and computational biology.