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: preregistered cross-domain replication study of PCA(50) effects on H1 persistent-homology signal across four omics cohorts, with confirmatory replications, ablation and confound analyses, a complete manuscript, and a citable Zenodo record.
- 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 preregistered evidence, calibration analysis, and explicit curation limits.
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.