Hospital Director
- Trying to
- Keep the hospital ahead of occupancy spikes
- Decides
- Beds, budget, extra staff
“We scramble when occupancy spikes, but the data arrives too late.”
Product DesignUX ResearchUI Design
Zonepulse
I designed a hospital operations dashboard, then let department heads choose its AI. They picked a capacity simulation over written summaries, and that version shipped.
Summaries vs capacity simulation, tested with department heads; the simulation shipped
Patient flow, staffing and compliance unified into one view
Director, ICU head and compliance officer — one dashboard, three lenses
Trend and scenario insights generated from live feeds, flagged before critical load
How might we …
Give hospital leaders one live picture they can plan from, hours before a surge instead of minutes into it?
Key screens
A hospital operations dashboard built around one question: what needs a decision in the next few hours? From the live command view to an AI insight and a tested plan.
The challenge
Zonepulse is a hospital operations dashboard I designed from zero for administrators and department heads. Patient flow, staffing and compliance lived in separate systems. By the time anyone pieced them together, the delay or the compliance risk had already happened.
Research
A department head sees the bottleneck forming in time to actually do something about it.
I interviewed people in three roles, audited every live data source, shadowed shifts on the floor, and tested early versions against real decisions. The same thing kept coming up: the hospital already had the data to see problems coming. It just never reached the person who could act, in time to act.
How a hospital sees a surge coming, and who acts on it
Before choosing what to build, I mapped how capacity decisions actually get made: who makes them, on which time horizon, and with which data.
How it works
Draft · confirm3systemsPatient flow, staffing and compliance each lived in its own system, refreshing on its own clock.
3horizonsThe director plans two weeks out, the ICU head acts on the next two hours, compliance reports after the fact.
1blind spotThe data to see a surge coming already existed. It never reached the person who could act, in time to act.
Decision horizonHours from now →
Who decides
Draft · confirmThree roles read the same hospital on different clocks. The quotes are from the interviews.
How far ahead each role planslog scale
ICU Head
- Trying to
- Never hit critical load unprepared
- Decides
- On-call nurses, admissions
“I need real-time alerts before we hit critical load.”
Compliance Officer
- Trying to
- Catch a risk before it becomes a filing
- Decides
- Escalations, staffing ratios
“Reporting is reactive. By the time it's filed, it's already a crisis.”
What already existed
Draft · confirm13%Best alternativeShare of the 4 capabilities the strongest existing tool covered.
100%ZonepulseBuilt to cover all 4, from seeing the hospital to testing a plan.
What leaders could useSee it all at onceWarns hours aheadSays what to doLets you test a planCoverage
The hospital's own systems0%
Tableau Healthcare13%
Epic dashboards13%
Zonepulse100%
Has it Partly (counts as half) Missing
Benchmarking Epic and Tableau Healthcare confirmed the gap. Centralised dashboards with predictive alerts were rare in Swedish systems: most showed what had already happened, not what was about to.
Why forecasting, of everything I could have built
Research surfaced five opportunities. I scored each on how hard users pulled for it, what the hospital needed, what we could build, and what nobody else offered. The top two became the product. The rest were built small, parked or gated.
Opportunities, ranked
Draft · confirmOpportunity score, out of 20UsersBusinessBuildGap
Dashed line: the lowest picked score (17). Everything at or above it went into the product.
What had to be true
Draft · confirm- A1Leaders will act on a forecast hours before a surge
- A2A written explanation is enough to act on
- A3One live view can serve three roles
- A4Showing the reasoning keeps trust through false alarms
- A5The feeds can refresh close enough to real time
- A6Staff will accept shifts the model pre-fills
Directions I explored
Draft · confirm3Directions explored
1Settled by an A/B test
1Kept as the structure
- 01dropped
AI-generated starter
Before opening Figma, I pressure-tested the structure with Gemini and Claude, then had v0 generate a starter dashboard to react to.
Result It made the wrong choices obvious fast: too dense, a flat hierarchy, charts built to impress. Which metric an ICU head needs at 7am came from context, not the model.
- 02picked
A lens per role
One data source, a tab per role, each tuned to what that role acts on.
Result One truth for everyone, without the director's view drowning the ICU head.
- 03tested
AI that explains vs AI you can test
A written summary of each prediction, against a capacity simulation leaders could move.
Result Settled by the A/B test: the simulation shipped.
What I built
Three decisions, each with a cost I chose to accept.
One live command view. Beds, staffing and equipment in one place, filterable by department, role and time window. The cost: integrating systems that refresh on different clocks.
Glanceable over dense. The wall has to read from across a ward under pressure, so status comes first and drill-down second. The cost: deep analysis moved off the main screen.
An AI layer that assists, never decides. It flags staffing and equipment gaps as they form and points to the units most at risk. The staffing and bed calls stay with the people accountable for them. The cost: forecasting hours ahead raises more false alarms than a reactive threshold, so every prediction has to show its reasoning.
Decisions backed by data
My first version of AI Insights explained each prediction in a written summary. I tested it with department heads against a version that let them simulate capacity themselves, and let the result decide what shipped.
What testing told us
Control vs. variation
A · ControlSummary-based insights
B · VariationCapacity simulation
What shipped: Version B
Every AI prediction now opens with its recommended actions next to a capacity simulation, so a department head can test the plan on the forecast before committing staff or beds.
2 leversAdmission and discharge rate, the numbers leaders already plan in
6 hoursOf projected ICU occupancy, redrawn as the levers move
3 actionsRecommended moves, each with its expected effect on shortage risk
Zonepulse: from a status wall to a plan you can test.
The whole hospital on one live screen, the AI insights one click behind it, and one alert followed across three tabs.
The whole hospital on one screen
Patients, beds, surgeries and staff load as one live view: bed occupancy by area, the alerts behind it, 30-day trends and coverage gaps. A director sees where pressure is building before anyone calls.
From summary to AI insights in one click
Generate Summary opens AI Insights over the overview. Every prediction is sorted by the decision it needs, with its confidence, where it came from and two next steps.
One alert, followed across three tabs
Analytics forecasts the ICU filling up, Departments shows the ward behind it with its staff and open shifts, and Quality & Safety shows whether care and reporting stay on track. Same live data, one tab apart.
What ships now, and what has to earn its place.
One live surface was feasible today. Letting the model act, not just advise, waits on trust it hasn't earned yet.
- Now
Plan on the forecast
A live view per role, with AI insights you can simulate before acting.
Live in V1 - Next
Forecast earns its trust
Confidence and reasoning on every call; the wrong-prediction state designed first.
Feasible now - Later
Predictive auto-staffing
The model pre-fills shift changes, once it has earned the trust.
Gated on model trust
What happened
Trust in the model was the whole product, and I still designed the happy path first.
What I'd do differently
Prototype the wrong-prediction state on day one, not patch it in late.
