• Build AI assistants that work on sensitive health data without exposing patient information.
• Turn slow, technical data tasks into natural-language tools researchers can run themselves.
• Train non-technical teams to adopt AI tools confidently and independently.
• Build reproducible, auditable research software with the controls and evidence regulated work requires.

Longer term, I work toward post-quantum cryptography readiness and crypto-agility for sensitive-data systems.

A fail-closed pipeline that turns raw clinical study data into something an AI assistant can safely query. Upload a ZIP of datasets, forms, and dictionaries, and the Judge classifies every column with one of three verdicts: keep it as-is, drop it, or apply a Safe Harbor transformation. The LLM only ever sees column headers, dictionary text, and form fields, never a patient row. A sandboxed executor applies the transformations, a deterministic verifier checks the result, and a residual PHI guard scans everything before it can publish. Anything it cannot classify with confidence holds the whole study rather than being silently dropped or guessed at.

• 11-stage documented pipeline, from intake through publish guard
• 340+ regression tests
• Source data is symlinked in, never copied or modified
• Documents its own known limitations, like a disclosed weak-fallback path, instead of hiding them

• QI Stat Studio: a guided statistics app that lets Rutgers IM Clinic medical residents run quality-improvement analyses on their own data, from upload to a shareable report, without needing a statistics background.
• Amarel VS Code: a VS Code Remote-SSH setup tool for Rutgers' Amarel HPC cluster, fixing the GLIBC 2.28 compatibility error on macOS and Windows.
• BatchQC: my first project, 72 commits to the R/Bioconductor package for batch-effect evaluation of high-throughput genomic data. NIH-funded, bioRxiv preprint 2026.

I work across Claude, OpenAI, Google, and local models, with RAG, prompt engineering, and evaluation harnesses under Model Context Protocol and LangGraph.