DayOne-HackML

Inspiration

In rural communities, prenatal care often comes down to a midwife, a phone and a pink paper booklet. Every visit is carefully written down, but the information never leaves the page. If the booklet is lost, damaged or left at home, the woman's history is gone. When she sees a new midwife, the story starts again from zero. And the health system can't learn anything from thousands of pages of good notes.

We wanted to keep the paper and the midwife's routine exactly as they are, and just let the information travel.

What it does

A midwife opens WhatsApp, photographs the pages of a patient's paper registry and sends them to an AI agent. The agent:

reads all the pages as one record and maps them to a fixed schema (medical history, obstetric history, antenatal visits, delivery, postpartum and newborn, vital signs, HIV, syphilis and hepatitis C tests)
gives every field a status and a confidence score (KNOWN, NEEDS_REVIEW, ILLEGIBLE, NOT_PROVIDED, NOT_APPLICABLE, UNKNOWN), so a blank is never just "N/A"
sends one short summary and asks only about the fields it's unsure of; she replies with a number: 1 Confirm, 2 Correct, 3 Retake photo, 4 Type it in
links the visit to the right woman using the random code the midwife writes on the booklet. It never creates a new patient when a match is plausible: she picks the patient, creates a new one, or answers "not sure", which sends it to a supervisor
writes de-identified rows to a shared Google Sheet and stores the original photo, unchanged, in a restricted Google Drive folder

No names, national IDs, husband's details, phone numbers or addresses are ever transcribed. Patient IDs (PAT-000001…) are sequential and never derived from personal data.

How we built it

WhatsApp + Muse (Meta's AI agent): the conversational front end. Midwives already use WhatsApp, so there's no new app, no new device and no training on a new interface.
One workflow skill file (Markdown): the agent's rules, schema, statuses, record lifecycle (CAPTURED → PENDING_AI → … → SYNCED, plus failure states), chat scripts, matching logic and roles. Every midwife's agent loads the same file, so everyone follows the same workflow. Changing the workflow means editing a text file, not rewriting code.
Google Sheets as the database: eight tabs (Patients, Records, Fields, Images, Review Queue, Audit Log, Codebook, Midwives) with a 71-field codebook of allowed values and plausible ranges. We chose Sheets over Excel because several midwives can write at the same time without overwriting each other.
Offline-first by design: with no signal, WhatsApp keeps the photo in its outgoing queue on the encrypted phone and sends it automatically when the connection returns. The agent uses the original message time as the capture time.
Roles: midwife, supervisor, IT admin and analyst, each with only the access its job needs. Analysts only ever see anonymous aggregates.
Dashboard: anonymized aggregates (blood pressure, temperature, infection testing) built from the synthetic dataset to show the public-health value of the data.

Challenges we ran into

WhatsApp Business Platform: activating it requires verified, legitimate business information, so we couldn't demo it.
Twilio WhatsApp Sandbox: the webhook connection failed in simulation, and even working, it offered document upload and processing rather than a real conversation.
Sijil, a standalone mobile web app: it needed a backend and hosted data storage, a recurring cost the initiative can't carry yet.
Privacy with no hosting budget: the hardest constraint. Our answer was to never write an identifier down in the first place, restrict access to original images, and be upfront that Meta and Google are third parties, so this version runs on synthetic data only.

Accomplishments that we're proud of

Zero change to the midwife's workflow: she photographs the booklet she already fills in, on the phone she already owns.
$0 to start: free tiers for Muse, Google and WhatsApp.
An agent that says when it's unsure instead of hiding it.
A workflow a non-developer can maintain: the whole behavior lives in one readable text file.
A complete implementation package: the skill file, a database template ready to import into Google Sheets, a setup guide with roles and permissions, and a printable bilingual command card for midwives.

What we learned

Uncertainty is data. Storing status and confidence per field is what makes human verification fast instead of a full re-check.
The best offline layer is often one users already trust. WhatsApp's own message queue solved a problem we were about to engineer from scratch.
Low-cost doesn't have to mean low-rigor. A clear lifecycle, an audit log and explicit roles fit inside a spreadsheet.

What's next

Growth tier: a $20/month Muse Power plan for the busiest midwives, plus Google Workspace admin controls once funding is secured.
Program tier: a self-hosted version (WhatsApp Business Platform, our Sijil backend and an in-country encrypted database) that fully meets the no-third-party rule and adds row-level access control.
On-device OCR so extraction can start without signal.
Automatically redacted copies of the original images for wider sharing.
A pilot with a small group of midwives once Muse is available in the region, or on the self-hosted tier.

NOTE: the DevPost includes a different that was create by Clause but was not tested with rigor. A demo is available though.