Dose Card
Turns a prescription into a large, picture-based medicine timetable that a grandparent can read on the fridge. It's printed in English plus their own language, from 15 options.
Built for Devpost's Build With AI: Basics hackathon with the Devpost Learn skill pack. The planning documents are in devpost/: scope, PRD, spec and the build checklist.
What it does
-
Upload a photo of the prescription, or press Try a sample prescription. Gemini's vision model copies out each medicine: name, amount, times of day and before/after food. It understands Indian shorthand like 1-0-1, OD, BD, HS, SOS, AC and PC, and marks anything unclear Please check.
-
Check each row against the paper and fix anything wrong, or type the medicines in by hand.
-
Pick the grandparent's language. Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Punjabi and Urdu are available, plus Spanish, French, Arabic, Chinese and Portuguese.
-
Print a one-page A4 tick chart: medicines down the side, Morning / Afternoon / Evening / Night across, with pictures for each time and for before or after food.
-
Scan the QR code on the card with any phone to hear it. Tap Morning, Afternoon, Evening or Night and the phone reads what to take, in the card's language, so it works for grandparents who can't read. Share on WhatsApp sends the same talking card to family. The card travels inside the link (after #), never to our server.
It never gives medical advice. It only copies the prescription, and the family member confirms every row.
Evidence
On 17 fictional test prescriptions (13 in different handwriting styles plus a typed discharge summary, and 3 hard cases: an elaborate script, a very blurry photo and instructions in Hindi), the reader:
- found 72 of 72 medicines;
- read 287 of 288 details right (amount, unit, when, food);
- made up none.
See the Evidence page. The test prescriptions are computer-drawn and cleaner than real handwriting, so they're a best case. Re-run with pnpm render-evals && pnpm eval; cached readings are in evals/raw/.
Run it
Needs Node 22+ and pnpm.
cp .env.example .env.local # add your Gemini API key (Anthropic key optional, used as a backup)
pnpm install
pnpm dev # http://localhost:3000
pnpm test # unit tests
Without a key, typing and printing still work, and the sample prescription uses its saved reading.
How it works
app/page.tsx + components/: the single screen with the start buttons, medicine list, card details and live card preview. The list is saved in this browser's localStorage only.app/api/read/route.ts + lib/reader.ts: the one server function. It sends the photo to Gemini (gemini-3.5-flash, falling back to two other Gemini models, then to Claude claude-sonnet-5-5 if an Anthropic key is set) with a fixed JSON schema, then validates the result with zod. The photo is never stored or logged.lib/share.ts, components/CardQR.tsx, app/listen/: the talking card. The QR link carries the card in its # fragment, and /listen reads it aloud with the browser's speech voices (lib/speech.ts).lib/translations.ts: the fixed card labels for each language. Switching language is instant and never calls the AI.app/card.css: the card and the print rules, which print only the card on A4.
Notes
- Sample data:
public/sample/prescription.jpg is a fictional prescription made for this project. public/sample/reading.json is a real reading of it, used only if the live reader is unavailable, and is labelled "Sample reading (saved)" in the app.
- Translations were machine-drafted. Corrections from native speakers are very welcome.
- Privacy: no accounts and no database. The photo is sent once to Gemini (or Claude, if every Gemini model is busy) for reading and isn't kept.
Licence
MIT
