I’m a scientist-turned-engineer
who keeps ending up at the
start of things.

Building the model and the framework to interpret it.

The product needed a survival model that would work across an unusually long horizon — three decades of reproductive aging, where most clinical models give up after five or ten years. I built it from scratch on longitudinal hormone and biomarker data, then spent as much time on what the model was for as on the model itself.

Interpretation is part of the model, not a step that comes after it. Clinicians and patients needed numbers they could act on — which meant translating a survival curve into a small set of metrics that mean something. I developed an ovarian-age framework from scratch: a peer-relative metric for where a patient sits among others her age, a gap metric that lets her track her own progress over time without being confounded by natural aging, and a zone system that translates a continuous prediction into clinically meaningful regions. None of that interpretation language existed in the field before — building it was as much of the work as building the model.

The report is the product.

A model in a notebook is not a clinical product. The work between those two things is mostly invisible — and it’s most of what decides whether anyone ever uses what you built.

The clinical report is the product. Everything I built around it — patient intake, provider portal, exports, the deployed model behind a versioned API — exists to put the right numbers in the right hands at the right moment in the workflow. The platform flexes across 13 user personas and clinical workflows because the report has to land cleanly regardless of who’s generating it and why.

Most of the past six months has been platform and product engineering — the kind of work that doesn’t show up in ML publications and rarely shows up in case studies, because there’s no clean experimental result to point at. But it’s the part where most clinical AI products actually fail: not because the model was wrong, but because the system around it never reached anyone who could use it.

LLM-powered clinical abstraction across the patient record.

As first technical hire, I designed and shipped a system that used LLMs to abstract complex clinical concepts from de-identified patient records — not just structured fields from a single note, but concepts that required reasoning across an entire patient's longitudinal history, sometimes hundreds of pages of clinical text.

The harder problems were accuracy and trust. I built guardrails to minimize hallucinations in the abstraction step, and a QA pipeline that improved accuracy iteratively rather than relying on the model to be right the first time.