WhiteCoat · Singapore
Role
Product Designer
Surface
Patient App · Conversational AI
Timeline
2025–2026
Team
PM · AI/ML Engineering · Clinical Advisor
Status
Built, pre-launch
The entry — opens from the home screen, before any service is chosen.
The final product — “I have a very bad flu.” Two questions, one route. Recorded from the build.
The hard case — “I’ve been feeling depressed.” Self-harm is the first question.
Where it started
Offered under the doctor’s card and the patient’s place in line. Useful preparation — but only for patients who had already picked GP.
Chest pain could be booked as a video call and only flagged mid-queue.
Specialists, mental wellness and chronic reviews never met it.
Its best answer was “wrong door”, to someone already in the doorway.
Where should an AI triage sit, if it’s meant to guide the choice rather than correct it?
What we learned
In the queue it only met patients who had already picked GP.
Options I considered
A form before booking, which I rejected because it decided who got care. The queue, which reached one service. The home screen, before any choice.
Why this approach
On the home screen, triage comes before every service — and every pathway is one sentence away.
BEFORE
Open app
→
Pick a service
→
Book a GP
→
Cody, while waiting
→
Consult
Triage starts after the choice — and only for GP.
NOW
Open app
→
Care Assistant
→
GP teleconsult
Emergency · 995
Specialist
Self-care
Physical visit
Mental wellness
Chronic care
Triage starts before any choice — for every pathway.
When the AI has heard enough, what should it actually say?
What we learned
A verdict closes the door on the patient. A menu of options hands the decision straight back to them.
Options I considered
A suitable / not-suitable verdict, a panel of four choices, or one clear recommendation.
Why this approach
A named next step, one action, and Back Home on every routine card.
Rejected
“Please seek in-person care”, and the only way out is Back to Home.
Superseded
Four open doors — kinder, but the patient still has to decide.
Chosen
1
The route, in plain words. The patient knows where they’re going, and why, before they see a button.
2
When not to wait. The signs that mean seeking help sooner.
3
The disclaimer. Guidance, not a diagnosis.
4
One action. It opens that service’s existing booking flow — nothing new to learn.
5
Back Home, then Chat Ended. The patient can always overrule it, and the chat closes instead of drifting.
Shown from the design file.
What happens on the day the AI is wrong?
What we learned
Including an emergency, in the middle of a conversation about something else.
Options I considered
Let the model judge urgency, or keep red-flag rules outside it. When unsure, guess the closest route, or send the patient to a GP.
Why this approach
The system is allowed to be wrong only in the safe direction.
Must never happen
An emergency sent to self-care — so it’s guarded by rules, not by conversation.
Accepted on purpose
A sore throat sent to a GP. Costs time, not safety.
Recoverable
Right urgency, wrong door. The chat stays on screen, so the patient can see why and choose otherwise.
One scenario, start to finish
I have chest pain.
Assistant
Says it’s an AI triage tool, then asks one question: any shortness of breath, sweating, or pain spreading to the arm, neck or jaw?
Shortness of breath.
Assistant
Asks how long, and whether it feels heavy, pressing or squeezing.
Suddenly, it happened today.
Red flag
“You are describing signs of a medical emergency. Please call 995 immediately.” One red button. No Back Home. Chat ended.
Recorded from the build, at 1.5× speed.
Consent
The first time a patient opens the assistant, one screen sets the terms in two lines: their answers are processed by AI to suggest a form of care, and the chat and a short summary are shared with their doctor. They can agree, or choose Not Now. Recommendations then end with the same reminder: an AI triage tool, guidance not a diagnosis.
Now playing: Emergency
Select a route to watch it
Emergency
“I have chest pain”
then: shortness of breath, started suddenly
Ends the chat on Call 995 — the only card with no Back Home.
995
Physical visit
“I have a very swollen cyst on my thigh”
then: redness around it, getting worse
An in-clinic GP nearby. A camera can’t press on a cyst.
IN PERSON
GP teleconsult
“I have a very bad flu”
then: no breathing difficulty, three days
On-demand video consult — and the fallback whenever it’s unsure.
VIDEO
Specialist
“My GP has referred me to see a dermatologist”
then: a few months, not painful, no fever
Only with a referral in hand. The same acne without one goes to GP.
REFERRAL
Mental wellness
“I have been feeling depressed recently”
first: a question about thoughts of self-harm
Safety is asked before anything else, then a psychologist.
PSYCHOLOGIST
Chronic care
“I have diabetes and would like to review my condition”
then: no new symptoms, taking metformin
A chronic disease review, after checking symptoms and medication.
REVIEW
Self-care
“I have a light headache, from two hours ago”
then: no fever, no head injury
Doctor-approved essentials, plus when to seek care instead.
AT HOME
FOUND IN TESTING — SAME CONDITION, NEARLY THE SAME SENTENCE, TWO ROUTES
I am diagnosed with diabetes and my doctor told me to have a follow-up consult to review my blood test
Route to: GP
I have diabetes, and would like to review my chronic condition with a doctor
Route to: Chronic Care Programme
A chronic patient shouldn’t land in a different service because they said “blood test” instead of “review”. The fix is a clarifying question, not a smarter guess.
What we can say
All seven routes run end to end in the build.
In testing, red-flag combinations reached 995 before any booking.
Every non-emergency route offers Back Home.
What isn’t proven yet
How real patients phrase symptoms, and how often routing is right at volume.
Whether patients accept the recommended route.
Clinical outcomes. Clearer routing is a usability result, not proof of safer care.
Limit
It’s the right urgency, but a crisis-line route should come before scaling that pathway.
Limit
A router that can’t admit confusion will guess.
Next steps
A clarifying question for chronic patients, so phrasing doesn’t decide the service.
A crisis-line route, distinct from 995.
An “I didn’t understand that” state.
A small task-based test with target patients: can they explain what to do next, and why?
Ranking the failures — below the need, above the need, wrong service — turned an open-ended safety debate into one rule: be wrong only in the safe direction. It showed which checks had to sit outside the model, and which mistakes we could accept.