Start with three products you can open. Then inspect the employer outcomes, where I name my role and keep the work unlinked because it is not mine to demo.
An independent consultant has a hunch and a deadline, and no cheap way to
find out whether the hunch survives contact with a hard question.
What I built
A decision accelerator that grills your plan the way a good advisor would,
then hands back a shareable decision record. Anti-sycophancy is a design
constraint, not a setting: a board of distinct lenses argues with you
rather than agreeing.
Key decision
The artifact is the product. Chat is the input; the shareable decision
record is the work product, which is what let the whole thing run for a
first-time visitor with no signup at all and turned the product into its
own marketing.
Stack
Next.js · Supabase · Anthropic API · open source
Healthcare strategy teams need US Census demographics to size a market, and
getting them means either learning SQL or waiting weeks on a consultant.
What I built
A natural-language interface over Census data. You ask in English and get a
sortable table and a chart back in under two seconds, across all 3,144 US
counties and 239,741 block groups on 84 variables, holding conversational
context so the second question builds on the first.
Key decision
An MCP validation layer between the model and the database instead of
trusting generated SQL: table and column allowlists, row caps, and
dangerous-pattern blocking all run before anything reaches DuckDB.
Stack
TypeScript · Next.js · DuckDB · PostgreSQL ·
Claude via the Agent SDK and MCP · v1 shipped February 2026
Architecture diagram
A diagram of the system, not a product screenshot. CensusChat has no hosted
demo — the source is the artifact, and this is where the gate sits.
It’s 9pm, something hurts, and you can’t tell whether this is an
urgent-care trip or a ride-it-out.
What I built
An AI symptom checker that routes to a next step rather than guessing a
diagnosis: self-care, go see someone today, or an async review by a real ER
physician who sends back a work note or a consult. Co-founded with Jon
Pangia, DO, a practicing emergency physician who owns the clinical workflow
and carries the risk on every paid deliverable; I own the architecture, the
eval pipeline, and the guardrails.
Key decision
A hard line at navigation with a clinician in the loop, and no crossing
it: not a diagnosis, not an AI doctor, not a medical device. Every deploy
is gated by a clinical eval pipeline with human physician review, because
in this category the thing that kills you is being right most of the time.
No links here. These shipped inside employers, and the products aren’t mine
to demo.
NeuroBlu Analytics
Director of Product, business-unit owner
~$10MARR
Real-world-evidence platform for mental health. Zero to ~$10M ARR. Scaled the
clinical dataset from 0.5M to 46M+ patient lives and 59M+ clinical notes.
Built the eval frameworks governing production GenAI across 25+ regulated
clinical organizations.
Revive Market Profiler
Sr. Director, Data Products & Analytics
$1MARR
Physician-referral graph product line. Grew the line to $1M ARR, up 225% over
four quarters, across 12 health-system clients.
HCA Patient Experience Dashboard
Data Product Manager
175hospitals
Enterprise patient-experience product. Deployed to 175 hospitals and 1,500
executive and nursing leaders. Cut report-development cost by $1.2M a year.
Trained 200+ users.
No links by design — employer work stays unlinked.
04 / Next step
Compare the record with the role.
If the work maps to the product problem you are hiring for, the next useful step is a direct conversation.
