I'm Paul Greenwood: data scientist and developer working on document ingestion, retrieval systems, and agent tooling, with a background in biomedical informatics and federal geospatial data.
Most of what I build lately answers the same question: how do you get messy, unstructured documents into a form a model can actually reason over? Scanned PDFs, OOXML, lab forms, literature, road inventories. The repos below are where I work that out.
Document ingestion & retrieval
- lab_submission_rag — RAG system that extracts laboratory submission data from PDF/DOCX into seven structured categories with validated models and per-field confidence scoring. ChromaDB, domain-specific chunking, pluggable Ollama/OpenAI/Claude backends.
- pdf_slurper — PDF table and metadata extraction (PyMuPDF + pdfplumber) with header normalization, numeric parsing, and content-hash idempotency into SQLite. CLI, FastAPI web UI, containerized.
- bb_rag — a RAG engine in Rust with no vector database and no framework. Sentence-aware chunking with overlap, TF-IDF and dense cosine retrieval, three swappable providers. Written to understand the internals rather than import them.
- LitKG — integrating biomedical literature with structured knowledge graphs (CIVIC, TCGA, CPTAC) for hypothesis generation. Entity linking with ontology disambiguation, hybrid GNN + RAG.
Agent tooling & Rust
- capscan — diff the capability surface of a crate between versions: unsafe, FFI, process/network/filesystem access, build scripts. On crates.io.
- capscan-mcp — MCP server exposing that scanning as agent-callable tools, so an agent can check what a dependency actually does before recommending a version bump. Ships with a Claude Code skill.
- decidex — extracts engineering decisions from git history and surfaces them in
CLAUDE.md, Cursor rules, and Copilot instructions, so AI tools remember what you already decided. On npm. - tracelet — minimal embeddable Rust tracer: the
#[tracing::instrument]you already use, without the OpenTelemetry dependency tree.
Science & geospatial
- ai-chipseq-mcp — applying ChIP-seq peak-calling methodology to transformer attention patterns for model interpretability.
- usda-forest-viz — downloading and visualizing USDA Forest Service Enterprise Data Warehouse geospatial datasets.
Data Scientist Leading Solutions LLC: Supporting USDA Forest Service Hurricane Helene recovery for the National Forests in North Carolina. OCR ingestion pipelines for thousands of scanned survey documents, Python/ML analysis of forest road infrastructure across 1,900+ segments tied to billions in recovery funding, ArcGIS Online automation, and long-term strategic planning analysis.
Product Manager & Developer UNC Chapel Hill School of Medicine: AI/ML research tooling, genomics pipelines (NCGenes), clinical NLP, and FHIR/HL7 integrations.
Project Manager UNC Chapel Hill: 20+ concurrent research projects.
Program Coordinator UNC Chapel Hill: EHR and precision analytics initiatives.
- AI/ML — RAG architecture, embeddings and retrieval evaluation, LangChain, ChromaDB, MCP servers, OCR (pytesseract), NLP, scikit-learn
- Languages — Python, Rust, TypeScript, SQL, R, Bash
- Data & geospatial — pandas, GeoPandas, ArcGIS Online/Pro, matplotlib, seaborn, Plotly
- Health informatics — FHIR, HL7, EHR systems, clinical data pipelines
- Infrastructure — Linux (Fedora, CachyOS), Docker, Slurm, Singularity, Conda, Git, CI/CD
- MPS, Biomedical Health Informatics — UNC Chapel Hill
- BA, English — UNC Chapel Hill
- Additional coursework — Chemistry (UNCG), Computer Science (University of Bergen), Linguistics (University of Oslo)
- Co-author, Nature — GENYSIS
- Wahl et al. (2021) — SARS-CoV-2 treatment research
English (native) • Norwegian • Swedish • Danish
Retrieval and document-processing systems • agent tooling and MCP • biomedical informatics • open-source research tools
Voracious reader • backcountry skiing • hiking • camping • cooking • acoustic guitar