Founding product hire at Vial, APM to VP in four years. Built the eClinical platform from zero — EDC, RTSM, portals — for patients, sites and monitors, then the AI tooling that runs on top of it. Before that: oncology research coordinator on CAR-T and immunotherapy trials at UCSF and MGH.
The 4pm Slack reply, the ticket, the tradeoff explained to a coordinator — that is what people experience of a PM. I write clearly, early, and to whoever has to act.
The question I ask most, and the easiest to avoid. If the honest answer is a shrug, I stop, say so, and redirect the effort.
Reason from what the trial, the site or the company actually needs — not from the roadmap, and not from what a vendor happens to do.
Follow the line. It starts with patients, runs through the preclinical bench and the clinical trial, and ends with the AI layer that now sits across all of it. Open any stop for the full case study.
As an oncology research coordinator I consented patients onto CAR-T and immunotherapy trials, chased queries at 7pm on Fridays, and learned exactly what a badly designed eCRF costs a site. I co-authored a molecular biology paper along the way (Jiang et al., 2016). Every product decision since has been a vote for the person entering the data.
Before a drug reaches a patient it generates thousands of assay rows nobody has time to clean.
Vendor cytotoxicity data that scientists hand-normalised became a validated pipeline with QC flags, a one-click IC50 curve generator and a 71-assay automation roadmap.
The call Flag, don't fix — the pipeline surfaces, the scientist decides.
The part of the pipeline where the software is regulated, the people are tired, and every field has a reason.
Led Vial's entire eClinical stack from a blank repo: five applications, 21 CFR Part 11 end to end, roughly 4.7× cheaper than vendors across five trials.
The call Build, not buy — but only what had to be ours.
Bi-weekly MedDRA/WHODrug coding, a query-review playbook and version-level eTMF tracking against the DIA model — pipelines instead of a data manager's chores.
The call Automate the chore, keep the human signature.
Validation packages signature-ready at merge, a documented hotfix path, and tickets specified tightly enough that a coding agent implements and one engineer reviews.
The call Slower merges, faster trials.
AI that touches trial data has to earn it. Mine keeps a human in the loop until it does.
An agent that reads EDC data and drafts monitoring findings. Catch rate ~60% → ~92% over six iterations; weekly review 8–12h → 2–3h; override rate steady at 11%.
The call Tuned for misses, not noise — decision support, not decisions.
Investigator's Brochures, Pre-IND packages and protocols drafted from the program database with a shared rubric and modality overlays — writers edit instead of starting from blank.
The call Drafts you can argue with, not documents you have to trust.
I don't hand problems to a backlog. I sit with the coordinators and monitors, then build — and most of what I build runs in production the same month.
I sit in the standup, read the queries, and ship the fix that week. The roadmap is written in the site's problems, not in a slide.
Regulated AI starts assistive. Override rate is a feature, and autonomy is something a system graduates into with evidence.
A ticket precise enough for an AI agent to implement is a ticket precise enough for anyone. One engineer reviews; the team's hours go to judgement.
The release isn't done until the package is signature-ready. Documentation that lags the code is a bug.
How it actually shows up in a day — not a strategy, a habit.
A skill has already posted the daily market briefing to Notion. I read it; I didn't write it.
Standup. Two tickets specified tightly enough that a coding agent implements them; one engineer reviews by lunch.
A question comes up about enrolment. I write the SQL myself in Metabase and answer it in the thread, with the chart.
The clinical monitor has drafted this week's findings from the EDC. I review the overrides, not the records.
Prototype a new payment-milestone flow in an afternoon with Claude. The screenshot goes to ops and design to decide, not to debate.
Draft the weekly readout from Linear and the deployment log with a skill. I edit the judgement, then ship it.
The same instinct — notice a thing that should exist, then make it — pointed at weddings, textiles, poetry and sound.
A gifting brand for Indian weddings and festivals — welcome bags first, Holi and Diwali next.
Bridesmaid and groomsman proposal cards and event welcome signs — mehndi, sangeet, haldi — designed to not look generated.
Silk headscarves. Monsoon-named.
A collection, written in the long shadow of Tagore and Rumi.
Experiments in letting a model hum and a human arrange.
Open to Director / Head / VP of Product roles in clinical-trial tech, eClinical, digital health and life-sciences tooling. Remote US, NYC or the Bay Area.