AI-native health · Zero-to-one · 2025-Present
Designing trust into an AI-native health product
A patient and a clinician need opposite things from the same model output - clarity and a confident next step for one, exposed nuance and uncertainty for the other. I designed both from a single inference, and turned the hard part into a reusable system for how the product should behave as confidence and stakes change.
Overview
Heal Sooner is an early-stage, AI-native health product that translates dense clinical signal into something two very different people can act on: a patient who needs clarity and confidence, and a clinician who needs enough exposed nuance to trust the model and act on it. This was zero-to-one work in a trust-sensitive domain, where the design decisions that mattered most weren't visual - they were about the posture of the model itself.
Problem & context
The same underlying inference had to do two opposite jobs. Give the patient the clinician's view and you overwhelm them with uncertainty they can't act on. Give the clinician the patient's view and you strip out exactly the nuance they need to trust - and safely override - the model. A single "neutral" surface would fail both. And because this is health, the cost of getting the model's posture wrong isn't just a worse experience; it's misplaced trust in either direction.
Why it mattered
Adoption of an AI health product lives or dies on calibrated trust. Patients disengage if the product feels either dismissive or alarming; clinicians reject any tool that asks them to accept a black-box output on high-stakes decisions. The product only works if the model knows when to be confident, when to hedge, and when to hand back to a human - and if the interface makes that posture legible to each user.
My role
I led the design of the AI experience and the system behind it - patient and clinician surfaces, and the underlying logic that decided how each should behave. Because the hard decisions were about model posture, I worked closely with the founder and the ML/engineering side on what the model could express (confidence, flagged inputs, uncertainty) and partnered on clinical input to make sure the clinician surface earned real trust rather than performing it.
Key decisions
One inference, two surfaces - not two products.
The patient surface hides calibration metadata and routes confidently to a single next action. The clinician surface exposes signal weight, flagged inputs, and fast override paths. Same data underneath; opposite postures on top.
Turn "how confident should the AI act?" into a system, not a per-screen guess.
I built a posture matrix - confidence on one axis, stakes on the other - that decides at runtime how assertive the UI should be: defer to a human, confirm with reasoning shown, suggest and allow retry, or auto-execute with an undo path. It made the model's behavior consistent and reviewable across the product instead of a series of one-off calls.
Design the model's voice deliberately, or the engineer will do it invisibly.
When to sound confident, when to hedge, when to step back - these are UX decisions with clinical consequences. Making them explicit and encoding them into the posture system meant the choices were owned by design and clinical judgment, not left to default.
Tradeoffs & constraints
Exposing uncertainty to clinicians risks overwhelming them; hiding it from patients risks false confidence - the posture system is the negotiated line between those failure modes, and it needed clinical input to place correctly. Building the matrix as shared infrastructure was slower than hard-coding each screen's behavior, but hard-coding would have produced an AI whose personality contradicted itself from one flow to the next. In a trust-sensitive domain, consistency of posture is itself a safety feature.
Outcomes
Currently in active development; some detail is under NDA. Happy to walk through the full story in a call.