dayone-whatsapp

Inspiration

In low-resource settings, midwives record every pregnancy in a paper registry. That information never follows the woman from one visit to the next, so we wanted to digitize it without changing a single step of their work.

What it does

A midwife photographs a registry page in a WhatsApp-style chat. AI turns it into a structured record with a status and a confidence score for each field, and shows its doubts first so the midwife only checks what matters. The pages are then linked to the right woman through the registry code, all offline-first with an encrypted local queue.

How we built it

We built it in Python (FastAPI, Pydantic, SQLite with encryption) with Claude Sonnet 5.5 for extraction. We also used AI (Claude Code) to help us design, code and test the project, while we made the design decisions and checked the results.

Challenges we ran into

Telling apart "empty", "illegible", "unknown" and "not applicable" for each field was harder than reading the handwriting itself. Making the AI's confidence actually point to its mistakes took several iterations.

What's next for dayone-whatsapp

Next steps: connect it to the real WhatsApp Business API, test it on blurry field photos and Arabic handwriting, and run the model locally so real patient data never leaves the device.