Computational scientist · AI engineer · Bioinformatics PhD
I build systems for data that is too large, noisy, or high-stakes to reason through by hand — public records, scientific literature, microbiome genomes, sequencing reads — turning it into searchable, auditable software and plain-English insight.
Large works
Searchable, entity-linked, citation-backed databases for massive public document dumps, built from OCR, redaction analysis, full-text search, image extraction, transcripts, and knowledge graphs.
1.4M+ documents, 2.9M+ pages, 190+ reports, PII-redacted database releases, and a public interface used for investigation and verification.
Evidence-grounded extraction of organoid culture protocols into structured records, preserving verbatim quotes, DOI provenance, grounding status, and protocol uncertainty.
582 papers, 25 organoid types, 5,458 grounded reagent records, KGX export, and a Datasette analytics API with dozens of protocol-intelligence endpoints.
High-performance microbiome tools for in silico PCR, degenerate primer search, and realistic PacBio HiFi amplicon sequencing simulation.
AmpliconHunter2 is a SIMD-accelerated C rewrite, 5.9× faster than the original, scaling in silico PCR to ~2.4M-genome databases (ICCABS).
A semi-autonomous research system for Claude Code with literature search, provenance tracking, evidence extraction, LaTeX synthesis, peer review, experiments, and self-improvement loops.
24 skills, specialized subagents, validation hooks, literature APIs, peer-review gates, and session artifacts designed for auditable scientific work.
Research software
The research work centers on high-resolution microbiome profiling: long-read amplicon sequencing, realistic simulation, scalable primer search, and statistical analysis of noisy biological datasets.
Resume (PDF)Tools and experiments
A dataset-agnostic deck.gl viewer for embedding clouds with binary buffers, lasso selection, filters, hover cards, thumbnails, and shareable state.
Runs in production at 1.3M+ points and 2.4M+ edges, without forcing each dataset into a bespoke UI.
Natural-language, end-to-end web testing driven by AI vision. A Set-of-Marks approach targets stable element IDs instead of brittle CSS selectors, explores autonomously, and writes tests in plain English.
Every run keeps an audit trail: screenshots, prompts, decisions, network, and console logs.
One script to run Qwen3 locally with Ollama, a browser chat UI, and optional Claude Code integration through a local API endpoint.
Practical local-model setup for offline experimentation and lower-cost AI workflows.
Throughline