Matria | Day 1 | CodeML

From paper to a trusted record

💚 Matria began with a simple question: how can a maternal registry become a useful digital record without making privacy, connectivity, or human judgment an afterthought?

We wanted to move beyond scanning paper into the cloud. Matria is designed for the real capture moment, where a midwife may be using a phone, working with an unreliable connection, and handling sensitive information that deserves careful treatment.

Built for one-tap capture

Matria is a privacy-first React and TypeScript PWA that turns a paper registry into a structured, reviewable record.

A midwife scans a QR code and is securely paired in seconds with a temporary session and generated work ID. Then the capture flow stays simple: open the camera, capture the page, and review it. Matria automatically detects the page, corrects its perspective, masks identifiers, and presents a full-screen preview with two clear choices: retake or send.

When offline, the app encrypts the capture in IndexedDB with a device-bound, non-extractable Web Crypto AES-GCM key. When connectivity returns, it resumes syncing automatically.

Local AI with a human in control

Pages securely travel through Tailscale Funnel to a local Mac, where the server performs a second alignment and masking pass. FastAPI and SQLite manage the record lifecycle, while GLM-OCR and Qwen run locally on Apple Silicon through MLX to extract structured information without relying on cloud AI.

Matria does more than read a page. It validates fields, detects uncertainty, links records carefully, and guides review through a chat-based workflow. Sent pages appear as masked image bubbles, and reviewers can confirm a field, correct it, request a retake, or ask for manual entry.

If the app cannot find a page or the server cannot safely process it, Matria does not guess. It offers a retake or lets the user enter the record by hand.

What we learned

The hardest problem was not OCR. It was designing a system that stays useful when the photo is imperfect, the connection disappears, or the AI is unsure.

We learned that responsible AI is a product decision. It means making the easiest path safe, keeping data local, making failures understandable, and giving people the final decision.

✨ Matria is an evaluation prototype built with synthetic data that brings together zero-setup mobile capture, encrypted offline storage, Tailscale networking, FastAPI, SQLite, and local multimodal AI into one human-centered workflow.