Research
Building the instruments
Two tracks, one habit: build the thing that measures before trusting what it measures.
Agents
Agent Infrastructure
Agent “skills” — packaged procedural knowledge an agent loads at runtime — are now everywhere, and almost none of the claims about them have been tested. My work here is the plumbing and the measurement: skill systems that actually run inside a harness, benchmarks that put capability and safety on the same axis, reward-integrity instrumentation for agents that game their own evaluation, and static analysis for the supply-chain problem that arrives once skills become shareable artifacts.
-
Benchmarking how well agent skills work across diverse tasks.
- Implemented the native skill system for Terminus2, so agents could load and run tools as modular packages.
- Developed the MHC layer task module and designed the ablation experiments that measure how much each skill actually helps.
- Built the analysis pipeline end to end — data collection through figures — and ran code review across the repository.
-
Evaluating capability and safety of LLM productivity agents in simulated workspaces.
- Simulates five productivity services — Gmail, Slack, Google Calendar, Google Docs, Google Drive — as one coherent workspace.
- 44 tasks that test capability and safety together, rather than treating safety as a separate afterthought benchmark.
-
BenchGuard
Model-backed instrumentation for reward integrity in LLM-agent evaluation infrastructure.
- Phase-aware static taint analysis exposes reward-hacking paths before a run; a runtime counterpart attributes concrete agent use from infrastructure-side evidence.
- BenchGuard Trajectories: a human-labeled corpus of 456 adjudicated trajectories drawn from 31,000+ public agent runs.
-
The evaluation environment and runtime powering benchmark suites — SkillsBench, AgentFuzzBench, FrontierPhysics — across 100+ containerized tasks.
-
Mapping the detection boundary of static analysis for malicious agent skills. Once skills are shareable artifacts they are a supply chain. This work charts where static analysis stops being able to tell a useful skill from a hostile one.
-
What Keeps Agent Skills from Being Reusable?
Evidence from 138,000
SKILL.mdfiles — an empirical look at what people actually write when they write a skill, and why so little of it transfers between agents. -
Co-organizer of the first workshop on the design, evaluation, and optimization of procedural knowledge for LLM agents.
Biomedical
Biomedical AI
Single-cell and spatial assays produce far more data than anyone can interpret by hand. I build models that turn those matrices into biology: cell-type annotation, tumour microenvironment structure, and drug–disease relationships read off a knowledge graph.
-
DG-scRNA
A graph-convolutional network with cluster visualization that predicts cell types from single-cell RNA-seq — 97.2% accuracy across more than 150,000 cells. First-author manuscript.
-
Thyroid tumour microenvironment
Integrated scRNA-seq with spatial transcriptomics across 25 thyroid cancer samples; identified 50+ ligand–receptor interactions and two previously undescribed cell subtypes.
-
A drug-repositioning web server: 15+ deep learning models over a knowledge graph of 5.9M edges and 107 relationship types, with visualizations that explain why each drug is recommended rather than just ranking them.
-
CRISPR gRNA on-target prediction
Benchmarked CNN, RNN, and Transformer architectures over 50,000+ curated guide sequences to establish which inductive bias the problem actually rewards.
-
Earlier wet-lab and computational work: proteomic profiling and genome-wide ChIP-seq mapping of O-GlcNAc chromatin-associated proteins, which uncovered an O-GlcNAc-regulated genotoxic stress response.
Elsewhere