I build the evidence and evaluation layer for AI in biology: experiments that generate hard biological data, benchmarks that measure what models actually know, and retrieval or post-training systems that make their limits explicit.

Evaluation

• grounding-atlas measures whether biological properties recoverable from model states are also expressed in model outputs across biological representations.
• verify-or-trust studies how a general model should trust, verify, or defer when using fallible perturbation-model outputs.
• causalatlas tests when perturbation foundation models outperform simple baselines and how assay cost and scientific objectives shape their useful deployment regimes.
• agentic-drug-discovery-system packages CTDBench v0.2, verifier contracts, and an audited retrospective discovery slice focused on sickle cell disease.
• LabCraft-Eval is an Inspect AI environment for evaluating tool-using agents on molecular microbiology protocols in a seeded laboratory simulator, with deterministic trajectory scoring.
• sci-agent-verification-cascade is a testbed for a failure in scientific AI handoffs, where a claim loses its source, uncertainty, or required verification as it passes between agents, and tests a claim receipt plus a runtime gate against it.
• SpaceOmicsBench is a multi-omics benchmark for spaceflight data: 21 ML tasks, 9 modalities, and a 100-question model evaluation.

Safety

• narrow-model-safety-eval asks whether narrow scientific models encode actionable biology, measuring ESM-2 and ProteinMPNN directly rather than testing whether a chat model refuses text.
• bio-overrefusal-v0.1 measures the opposite failure with 201 domain-expert-authored, tier-annotated research queries, so the cost of refusing legitimate work is quantified alongside the cost of answering.
• llm-sfm-safety-eval evaluates refusal calibration when a general model interprets science foundation model outputs, across 24.3K outcome records.
• bio-sfm-trust-audit audits trust routing where a general LLM sits above specialist protein, genome, and single-cell models and has to decide what to accept.
• protein-label-integrity-eval tests whether a model notices a hazardous sequence stored under a benign-looking label, across five Claude versions.
• constitutional-bioguard is a leakage-clean harness for biosafety guard models and an honest case study: the guard trained here is Pareto-dominated by a smaller open model, and the repository says so.

Negative evidence

• NegBioDB / NullAtlas is a private, release-gated biomedical negative-results research program supported by Anthropic's AI for Science program. Its public negbiodb-safety-calibration companion reports aggregate, non-record-level calibration results across five biomedical domains.

Post-training

• I build and diagnose SFT, DPO, GRPO, and QLoRA workflows for biological evaluation tasks. A recurring result is that some evidence is better supplied through retrieval and deterministic checks than learned through weight updates.
• A bounded public example, BioReview_Training, publishes QLoRA training code and committed held-out v3 summaries; the 8B+9B ensemble reaches an F1 score of 0.704.

At Weill Cornell Medicine, I lead spaceflight research and co-founded the SOMA consortium, which spans more than 100 institutions across 25+ countries. I came to AI evaluation from drug discovery, single-cell genomics, and Perturb-seq. I hold a Ph.D. from KAIST and am an inventor on two licensed patents.