Off the shelf or tailored, evaluated and deployed for regulated settings. From a working prototype in days to AI woven across the build — with a human gate on every step.
A 30-minute call to get a first read on where AI earns its place in your product, and a proposal for a 2–4 week problem-framing engagement.
AI that earns its place in the product — proven in a prototype before it ships, gated by a human, and built to hold up under regulation.
Most AI in product fails not in the model but in the deployment — the wrong pattern, no evals, no human gate, no path through compliance. I work inside your team and your code to put AI where it moves a real metric, and to change how the team builds so the speed lasts.
Off the shelf or tailored. The right model and pattern for the job — copilots, scribes, summarization, retrieval, agents — placed where they move a real metric, not where they demo well.
Vendor and model evaluation, evals and guardrails, graduated autonomy. Deployed where work has to stay separated, compliant, and auditable — clinical, payer, and other regulated workflows.
AI across the build, not just in the product. How your team specs, prototypes, ships, and reviews — so the build itself gets faster and the org learns to work this way.
Not a deck. AI already deployed in a real clinical product, a system that runs my own practice every day, and a regulated product prototyped with AI in days.
The central proof: as SVP Product & Technology (CPTO) at Form Health (obesity and cardiometabolic care), I put AI directly into the clinical workflow — record review, scribe, summaries, and inbound-call handling — deployed and running in a live, regulated setting, not a pilot.
Every context in its own room, firewalled, while the learning compounds under your rules — agents, memory, guardrails, and audit as one system, built verify-not-promise so nothing ships on a promise. It runs my own portfolio daily, and is offered within engagements: deployed in your code, hosted, or managed.
Clinical programs were running on slow homegrown software. I chose a custom EMR on Medplum + Claude Code over extending the legacy stack, and used AI to prototype working capabilities in days, not quarters — a working prototype that proved a viable path off the old software, in a real clinical setting.
Most teams don’t need another framework. They need the specifics of how you make an unpredictable model safe and useful inside a clinical product. Here is how I made these calls myself at Form Health and Maven Clinic, and taught my teams to make them too.
You don’t ship a clinical AI feature because a demo looked good. Before anything went into Form’s workflow, I stood up an eval harness: a graded set of real, de-identified cases, scored against a rubric my clinicians wrote, and re-run on every model and prompt change. The feature ships only when it clears the bar, and a human still reviews the hard ones. That is how you turn “it seems to work” into a number you can defend to a CMO and an auditor.
The most important design decision is what you let the model touch. I use AI for the language-shaped work a human reviews anyway: summarizing a chart, drafting a visit note, triaging an inbox, handling an inbound call. Anything that has to be exact and provable, like eligibility, dosing, or billing, stays in plain, testable code.
My rule: the model sits behind the gate, never in front of it.
Each context and tenant gets its own walled room, so data never bleeds across. The model sees the minimum it needs: de-identified, scoped, pulled in just in time, with every access logged. Inference runs in-region, hosted or self-hosted, under a BAA where the data demands it. It is the same walled-room pattern behind mnemur.ai, pointed at clinical data.
Get these three right and AI stops being a demo risk. It becomes something you can actually deploy: separated, compliant, auditable, and gated by a human at every step.
Most engagements start with a 2–4 week problem-framing sprint, priced by role and cadence, not hours. The phases overlap.
Where AI actually earns its place, build vs. buy, which model for which job — with evals and a compliance path defined up front, not after.
A real capability in your product, not a demo — prototyped fast with AI, tested against clinical and operational reality.
Guardrails, graduated autonomy, and a human gate for the regulated setting — and your team building this way after I’m gone.
See the one-page capabilities summary →