The problem
A hypertension patient in Riyadh takes the same three drugs every month. This month one box costs more at the usual pharmacy; another is out of stock. The family calls four pharmacies, or just pays extra, or the dose gets skipped for a week. The information exists — stock and price per pharmacy — but nobody aggregates it, so every refill is a small scavenger hunt.
The build
Two people ship a WhatsApp number for refills. The patient sends a photo of the prescription or the empty box. A vision model reads the drug names, doses, and quantities, checks a stock-and-price table for pharmacies in the patient’s district, and proposes the cheapest one with everything in stock. One reply confirms; the pharmacy delivers or holds the bag for pickup. Next month the same refill re-triggers with a single confirmation message — no new photo needed.
Why now
Reading a prescription photo used to need a custom OCR pipeline and a pharmacy IT integration. Now a Flash-class vision model does the extraction at roughly thirty cents per million input tokens with a free tier, Maps grounding finds nearby pharmacies with hundreds of free lookups a day, and the confirmation chat back-and-forth on WhatsApp costs nothing inside the 24-hour window. The whole product is three APIs and one table.
First customer
Independent pharmacies in Riyadh residential districts. A routed chronic patient refills monthly for years — lifetime value no flyer campaign can match. They pay a monthly fee per confirmed refill because the economics are legible from the first order, and the pilot starts with the five pharmacies around one clinic cluster.
Challenges
A misread drug name is a safety failure, not a typo — every extraction passes a pharmacist review step until accuracy is proven, and controlled substances are excluded by policy from day one. Stock tables rot within days, so the pharmacist-facing update step must take thirty seconds or it dies. And prices differ by city and supplier, so the table is per-district truth, never model memory.