Oncology companion

The client brought an AI-generated dashboard and an MVP document. They pointed in different directions: the dashboard used streaks, progress, scores and strengths ratings; the MVP said the product should not be gamified.

I treated the MVP as the stronger constraint, so those elements stayed out of the product. Missing a few days should not feel like failure.

The early font also felt wrong, so I pushed back and moved the interface to Satoshi, which is calmer and easier to read.

streakprogressscore

fig. 01 — Client reference. Useful as a reference, not a product direction.

• Not in the MVP
• Missed days should not feel like failure
• Support, not performance

too much like an experiment

more space, calmer type, clearer return point

The product had four roles: patient, caregiver, coach and clinician.

These were working assumptions based on the MVP, seeded data and review notes, not research findings.

Personas and information architecture helped define what each role needed to do, what they could see and where the system needed to hand off.

This was not one app with permissions added later. The roles shaped the structure from the beginning.

Onboarding took longest, because the team needed to decide whether to personalise first or ask for an email address first.

I built both orders in the prototype so the team could walk them side by side: a conversation before the account, or the account and the promises first, with the personal questions moved inside the app.

I mapped the patient flows, including their exits.

A check-in could continue in the product, move to the care team or route to crisis support.

After writing in the journal, the patient decides who gets a summary: the page stays private, or they pick the sections and who receives them.

Behind the interface, the belief system updates after each interaction and uses clinician-created content to shape later responses.

I designed the visible check-in, recommendation and handoff states. The belief engine, classifiers and clinical operations were outside my scope.

Journal pages and check-in words stay private.

Caregivers, coaches and clinicians see different amounts of information, depending on role and permission.

The patient decides what can be shared before someone is invited.

This became a minimum-disclosure approach: show each person what they need without turning private experience into a shared dashboard.

I moved into Claude Code because of the deadline, but the layouts became repetitive while the flow was still changing.

I returned to Figma to settle the decisions, then rebuilt the direction in React.

The prototype was shared through GitHub with the full-stack developers, AI engineer, CTO and product manager.

It holds seven connected support flows, 21 resilience beliefs and three safety and escalation tiers.

The product is in beta. There are no health-outcome or clinical-impact results to claim yet.

These are the screens in the prototype today, on the app’s own night-tree art, phone and desktop.

It opens with three promises and an optional minute of quiet, so the first thing asked of a person isn’t a form.

The journal takes whatever you say, as it is. When an entry ends it is read back gently, sometimes with a short poem, never with advice. You can ask it what you wrote.

Asking the companion gets a calm answer, and sometimes a short video made for that moment.

Who sees what is decided per person, one switch at a time. The journal is never one of the switches, and everything the companion remembers can be taken off.