About the project

In many low-resource settings, maternal care is recorded in handwritten paper registers. Those records are vital, but they are difficult to carry across visits, search, analyse, or safely share. We were inspired to build a tool that respects the midwife’s existing workflow: paper remains the source of truth, while a simple photo becomes a structured, longitudinal digital record.

We built an offline-first, WhatsApp-like prototype for digitising maternal registers. The pipeline preprocesses difficult images, detects text regions, recognises handwritten French and English with a dedicated handwriting model, detects checked form boxes, and maps extracted labels to a clinical registry schema using multilingual semantic matching. Each field is validated against appropriate types and ranges—for example, maternal weight, blood pressure, gestational age, and newborn measurements.

The structured records are stored locally in an encrypted SQLite database. We chose append-only observations so corrections preserve history rather than silently overwriting earlier information. The chat interface supports reviewing and correcting stored values, while the data model supports explicit uncertainty states such as CONNU, ILLISIBLE, and À_RÉVISER.

One major challenge was unreliable input: blurred, tilted, poorly lit images; mixed printed and handwritten text; medical abbreviations; and incomplete fields. We addressed this with image preprocessing, separate detection and handwriting-recognition stages, multilingual matching, confidence thresholds, and conservative validation. Another challenge was operating on limited hardware, so models load lazily, use GPU acceleration when available, and reduce batch sizes when memory is constrained.

This project taught us that digitising healthcare records is not just an OCR problem. Reliability, uncertainty, privacy, local usability, and preserving the clinician’s judgment are equally important. Our goal is not to replace midwives, but to give them a safer, faster bridge from paper records to continuous care.