I was Product Designer at minoHealth AI Labs from November 2023 to May 2025 (about 18 months). I owned product design for Moremi AI — the clinical interface, the uncertainty model, and the design system — not the diagnostic model and not the company's commercial results.
The design problem
Clinicians think in patients and diagnoses, not models and algorithms. The first versions of the product were organized the other way around.
Before: Models → Algorithms → Results After: Patients → Diagnose → Review → Report
That shift came from 23 clinician interviews and 15 contextual inquiries across Accra, Lagos, and Nairobi. It is the most important IA decision I made on the project.
The second problem was trust. In healthcare, uncertainty is information. If the UI only shows a diagnosis, it is lying by omission.
Design work
1. Clinical-first information architecture
I restructured navigation and the primary dashboard around the clinical loop: open a patient, read the study, review the model, write the report. Technical model controls moved out of the default path.
2. Designing for AI uncertainty
- Confidence first. Research said the first question was always "How confident is the AI?" — so the score sits at the top of the diagnostic view.
- Low-confidence mode. Below 80% confidence, the interface shifts to review: amber indicators, specialist consult prompted before proceed.
- Traffic-light states. Green / amber / red — readable in a time-critical glance.
- Human override. Every recommendation has a "Clinician disagrees" action. The model augments; it does not replace.
3. Offline-first, phone-first
The platform had to work in a teaching hospital and a rural clinic on 3G. I designed skeleton loading that mirrors the three-stage response (image processing → inference → confidence), offline sync for the diagnostic core, and layouts that hold up on a phone.
4. Multimodal diagnostic surface
One dashboard for imaging, patient context, and AI analysis: side-by-side original vs annotated image, heatmaps for focus areas, plain-language explanations.
5. Report generation as a designed workflow
I designed the report flow as a human-reviewed template system — generate, edit, export — not a dump of model text. Multi-language templates (English, Swahili, French, Yoruba) and print-friendly PDF were part of the same handoff.
Report generation as a human-reviewed workflow. Placeholder copy on the unused report pane was cropped out.
6. Design system
I built a modular component library in Figma targeted at WCAG 2.2, then translated it into a production React/TypeScript library so design and engineering shared states, not screenshots.
About the company and product
The figures below are minoHealth / Moremi product and company context. They are not my personal results. I designed the interface those clinicians used; I did not train the model or close the revenue.
| Context | Figure | Whose |
|---|---|---|
| Diagnostic performance (pleural effusion) | 97% AUC-ROC (n=2,400) | Company / model |
| Cardiomegaly detection | 90% vs 77–87% individual radiologists | Company / model |
| Report time (product claim) | 45 min → 3 min | Company / product |
| Facilities at launch (2024) | 500+ across Ghana, Nigeria, Kenya | Company |
| Platform revenue (2024) | $2.9M annual | Company |
| Later footprint | 50+ countries; 30 imaging modalities | Company, after my tenure |
| Media | CNN coverage of the platform; CEO Darlington Akogo named to Forbes 30 Under 30 (Class of 2025) | Company / CEO — not my award |
| Publications / M&A interest | Peer-reviewed papers; M&A offer (April 2025) | Company |
Ghana had fewer than 80 radiologists for 33 million people. That constraint shaped the design: the UI had to work for a specialist and for a rural nurse, on a phone, when the network dropped.
The design system I left is still the foundation those later markets sit on. That is the part of the scale story that is mine: the IA, the uncertainty states, and the components — not the AUC, the revenue, or the press.
What I learned
- Uncertainty is a first-class UI state. Designing the low-confidence path mattered more than polishing the happy path.
- Clinical IA beats model IA. If the nav says "Models," you have already lost the clinician.
- Offline is a Day-1 design problem. You cannot bolt connectivity onto a desktop-only mental model.
Related work: Moremi Collaborate — the peer-review and project-collaboration surface on the same platform. Varve — the design-engineering work I do now, from research and Figma to production React and TypeScript.