2,000 handwritten records, one weekend build.

The challenge. An international medical nonprofit was closing out a field clinic when its planned data-collection system became unavailable. Every patient visit and every dispensed medication had been written by hand into spiral-bound registers — more than two thousand records that existed nowhere else. The volunteers who understood those registers were about to leave.

Why the usual answers failed. Typing it all in was the only fallback: at one to two minutes a record, thirty to sixty-plus hours of data entry — by volunteers who could give two hours a day and weren’t fast typists. Weeks of delay, at best. Scanning apps assume reliable internet, tidy handwriting, and one page layout. This clinic had intermittent power, several register formats, handwriting in more than one language, and no budget for recurring services.

The stakes. If the registers left with the volunteers, the data was gone for a month or two — or for good.

What we built. An installable, offline-first mobile app. Photograph a register page and it is saved on the phone before anything else happens. When a connection appears, pages sync to the nonprofit’s own private Google Drive and Sheets. A server-side AI reader then transcribes every page into structured rows — reading each page twice with differently worded instructions and flagging any field where the two readings disagree. Volunteers check flagged fields against the photograph on screen.

The workflow. Open the home-screen icon → confirm who you are and which register → one photo per page → keep going. Later, anyone on the team opens To check and confirms the machine’s reading against the photo. At end of shift, the supervisor runs a reconciliation that proves nothing was lost.

Decisions that mattered.

• Save first, sync later. One atomic save per page; the volunteer is back at the camera in seconds.
• Confidence is observed, not asked. We never trust a model to report its own uncertainty.
• Nothing is silently dropped. Struck-through and illegible entries are transcribed and flagged — after a real case where “ignore crossed-out lines” quietly lost two dispensed medications.
• Privacy by default. Location metadata is stripped from every photo; phones are guided to keep patient photos out of personal cloud backups.

Timeline. Requirements captured in a 45-minute conversation. A complete working product existed 3 hours 18 minutes after the first commit. The field-ready system — training card, QR install, AI reader, review workflow — at 14 and a half hours.

Outcomes. The team reports more than 2,000 records digitized through the app, with digitization complete within a couple of hours of connectivity instead of weeks of typing. 148 register pages were captured and preserved with zero loss, proven by reconciliation. Measured cost: about $0.004 and 10 seconds per page. An eleven-day sync bug was later found and root-caused in a written postmortem — with no data lost, because the save-first design held.

What this proves. Paper-to-data rescue under field conditions: offline capture, a near-zero-cost backend on tools the organization already owns, and AI verification you can audit. The same pattern fits intake forms, logbooks, attendance sheets, and receipts.