LLM & agentic systems for clinical genomics

At Franklin by QIAGEN (I joined Genoox in January 2025; the company was acquired by QIAGEN during my tenure), I develop algorithmic solutions for complex genomic and clinical data — and lead the organization’s LLM and agentic-workflow initiatives.

What this looks like in practice

• LLM applications, end-to-end. I take LLM-powered features from “is this even feasible?” through prototype, evaluation, integration, and production deployment — in a domain where wrong answers have clinical consequences, so evaluation, testing, and reproducibility are first-class engineering artifacts, not afterthoughts.
• Agentic workflows. Designing multi-step, tool-using workflows that automate genuinely hard knowledge work in genomic interpretation, with the observability needed to trust them in production.
• Literature-scale NLP. Scientific literature is the raw material of variant interpretation. I build NLP and text-mining systems that turn papers and unstructured clinical content into structured, actionable evidence.
• Advisory role. I advise teams across the organization on AI solution design, technical feasibility, and technology selection — the internal “go-to” for AI development.
• Performance engineering. Optimizing data-intensive algorithms with parallel and distributed computing.

Why it’s hard (and interesting)

Clinical genomics is an unforgiving environment for AI: the source material is dense scientific prose, the ontology is huge, and the cost of a hallucination is not a bad demo — it’s a wrong call on a patient’s variant. Building LLM systems that earn trust here means treating evals as the product: measurable accuracy against expert-curated ground truth, regression suites for prompts and models, and graceful degradation when the model is uncertain.

The specifics of what we ship are Franklin’s story to tell — this page describes my role and the engineering approach.