I build provenance-first scientific AI systems: MCP servers, bioinformatics agents, and reproducible ML/research tooling.
My strongest public work is in agentic research infrastructure: tools that let AI systems query scientific sources with typed APIs, provenance records, verification hooks, and explicit limitations.
- uniprot-mcp — UniProt MCP server with per-response SHA-256 provenance, release pinning, verification, and offline replay.
- semantic-scholar-mcp — Semantic Scholar MCP server with 14 typed tools for paper search, citation graphs, author profiles, and recommendations.
- alphafold-sovereign-mcp — local-first biomedical MCP over AlphaFold DB and related public sources, with a SQLite knowledge graph and explicit clinical-use limits.
- pca-topology-crossdomain-replication: research on PCA(50) and H1 persistence across omics cohorts, with a corrected matched-null comparison, explicit methodological corrections, executable audit records, and versioned Zenodo archives. The matched study supports dependence-sensitive geometry, not an established biological topology beyond covariance; independent-cohort validation remains a separate research task.
- TopoGeoML — topology-aware ML toolkit plus a preregistered graph-classification study whose headline result is negative.
- cohomology-wall — mathematical/reproducibility archive for a tetraquadric Calabi-Yau cohomology-jumping calculation.
- GROUPOID — pre-alpha groupoid/sheaf/Riemannian aggregation prototype for federated learning.
- homology-cliff — ESM-2 protein-retrieval failure-mode study with archived experimental evidence, calibration analysis, and explicit limits on preregistration chronology, label curation, and external generalization.
The repositories below are retained as historical supporting material and are archived rather than actively maintained.
- democafa_package — archived CAFA-style protein-function prediction data and evaluation utilities.
- LAFA_container_guide — archived containerization guide for protein-function prediction methods.
- icbo-ai-tutorial — archived tutorial/codespace material for ontology-agent workflows.
I lead these projects independently and am the sole scientific author of their research contributions. AI-assisted tools support parts of implementation, analysis workflows, and internal review; research design, interpretation, and release decisions remain my responsibility. Upstream software, data, and prior research are credited in the respective repositories.