Talk is cheap. Here's what I've actually built.

The problem

Patients dread intake. You fill out the clipboard, hand it over, and hope the time you just spent adds up to something — you never really know where the data goes or what it's for. And what comes back is a label, not an understanding. Two people arrive with the same wound: the adult who's been told their whole life they're lazy, and the parent fighting to prove their kid isn't a behavior problem. A box of checkboxes gives them a category that lands as judgment. They leave knowing what they're called, not what's happening inside them. And for the clinic, every intake is hours of manual work that teaches the patient nothing before they even walk in.

The approach

Use AI only where it's genuinely better — and keep it out of every place a health decision can't afford a guess.

The clinical scoring is plain, auditable code — no model deciding anything that matters. The AI is fenced to the two jobs it's actually best at: reading a real, messy lab report, and turning cold clinical facts into language a person can feel. A human confirms every number before it counts. In health, knowing where to keep the model out is the whole game.

What I built

• ✓A humane intake — 52 questions in plain, human language, no jargon and no checkboxes that feel like a test, with the tone set by whether an adult or a parent is answering.
• ✓A lab reader — the patient uploads their actual lab PDF, and the AI pulls every marker, maps each to a plain-English name and an optimal range, ready for the person to confirm.
• ✓A report that mirrors the body back — an archetype, a body map, and the top overloads, written so a person sees themselves in their own data instead of a diagnosis they can't parse.
• ✓A clinician's page — the same data, print-ready, so the provider reviews the whole picture in a few minutes instead of the full manual workup.
• ✓A health journey over time — every intake and lab result tracked visit over visit, so the patient can see the positive changes as they happen.

What it proves

Why this is rare

The hard part isn't adding AI — it's knowing where to keep it out. Most consultants would wrap a chatbot around a health intake and call it innovation. In a clinical setting, that's how you get confident, wrong answers. I built the opposite: AI fenced to what it's genuinely good at, plain auditable code for every decision that matters, and a human in command.

Under the hood

The problem

One chatbot can't hold a whole person. Most AI products are a thin layer over a single chat box — one model, one voice, one short memory. Ask it to truly know you and it flatters you, forgets what you told it last week, and quietly blurs what you said with what it guessed. And the real work — pulling together your experiences, your values, your goals, and the patterns running through your own history — would take years to gather in one place and actually make sense of. Building something that can genuinely reflect, coach, and challenge a person takes more than a bigger prompt. It takes a real system.

The approach

Don't train a model on a person. Build an owned record of them — and bring the right slice to the model at the right moment.

The model stays smart, current, and swappable. The data — a structured, versioned record of one human — is the asset that keeps appreciating. Over it sits a board of seven specialized agents: the same frontier model given different roles. An orchestrator routes each turn to the right ones, runs them in parallel, and — for the hard questions — convenes a council that keeps the disagreements instead of averaging them into bland advice.

What I built

• ✓A board of seven specialists behind one voice — the Companion you talk to, a Biographer that interviews and records, a Challenger that argues against you using your own values, a Pattern Mirror that surfaces recurring behavior, a Council that holds opposing views on the hard questions, a Lookout that nudges and flags drift, and a Presenter that renders it all in the right visual form — run in parallel so you only ever feel one continuous presence.
• ✓A partnered process through every part of your life — it walks you through your experiences, values, goals, and history one thread at a time, and surfaces the patterns you're too close to see on your own.
• ✓One engine, a private version for every user — the engine stays impersonal and reusable; only your data is personal, so a deeply personal tool and a product for thousands of people are the exact same codebase.
• ✓A companion that learns as you go — every reflection is written back into what it knows about you, and newer insights replace the ones they outgrow, so its picture of you stays current without ever retraining the model.

What it proves

Why this is rare

Give me a vague, high-stakes ambition and I turn it into clear requirements, a defensible architecture, and a working system — engineered from day one to serve thousands, survive audits, and outlast the next model upgrade. That's the difference between a pilot that strands you and a platform you can build a business on.

Under the hood