Kuwaiti healthcare group · AI practice
A proof of concept: handwritten dental order forms, read by AI and confirmed by a person.
- 4 daysfrom written spec to the full order flow
- 3AI vision models read each form in parallel
- 2languages on patient pages, Arabic first
- 63commits
Patients bring in, or send photos of, paper order forms that doctors fill in by hand for crowns, bridges and implants. Staff re-keyed every field and sent payment requests as they went, and patients couldn't see where their order was.
That meant double entry, transcription errors, no audit trail, no shared view of status and no structured data for reporting.
- Upload with a readability check
- Patients photograph the form on their phone. The image is compressed in the browser, and a quick AI check flags a wrong or unreadable document.
- Review beside the original
- Staff see the photo next to an editable order, with uncertain fields highlighted, a tooth chart and the models' readings side by side.
- Payment links
- Confirming an order creates a KNET or card payment link. Payment status is re-checked with the provider before it counts. Payments run in test mode in the proof of concept.
- Order tracking for patients
- Patients enter a reference and phone number to see a plain-language timeline, in Arabic or English.
- Lab production and materials
- Orders move through production stages with appointment dates, and an internal workflow handles material requests.
Where AI works in the solution
Three models read each form. A person decides.
Claude, OpenAI and Gemini read each form in parallel, using one shared prompt that covers the form's variants, the tooth chart and Arabic-Indic digits. One model pre-fills the review, and a failed civil-ID check digit lowers its confidence. If one model fails, the others still produce an order to review. A person always confirms before payment is requested.
Models compared on real corrections
The app tracks each model's accuracy, cost and response time, scored only against corrections a person has checked, so the choice of model rests on evidence rather than impressions.
From a phone photo to a paid lab order
- existing systemPatient's phonea photo of the paper form
- AI modelsClaude · OpenAI · Geminiread the form in parallel
- we builtStaff reviewconfirm, price, track
- existing systemPaymentsKNET and card
| System | Direction | Method | What flows |
|---|---|---|---|
| Anthropic Claude | out | SDK | Form images in; order fields and the readability check out |
| OpenAI | out | SDK | Form images in, order fields out |
| Google Gemini | out | SDK | Form images in, order fields out |
| Payment gateway (KNET and card) | both | REST · webhook | Charge creation and payment status, re-checked before it counts |
| out | Click-to-chat link | Staff send the payment link from a prepared message | |
| Private file storage | out | SDK | Original and working copies of each form |