AI document triage in Pakistani healthcare that refuses to guess

1 The problem

A healthcare network was receiving thousands of referral letters a week as scans, faxes and email PDFs. Three people spent their days reading each one, picking specialty and urgency, and typing it into the patient system. Five-day backlogs were normal. An urgent letter could sit unread over a weekend.

2 Our approach

The design that got it past clinical governance was not the model. It was the threshold. We classify specialty and urgency; anything the model is not confident about goes to a person. Clinicians see extracted fields next to the original page and correct in one click. Those corrections are the next week’s labelled set. Built and monitored from Islamabad, not as a demo that only works on clean English PDFs.

3 What we built

• OCR for scans, faxes and native PDFs
• Specialty and urgency with a calibrated confidence cut
• Low-confidence cases to a human queue by design
• Side-by-side review with one-click correction
• Audit trail of model output and every override
• Weekly evaluation on a clinician-labelled holdout set

4 Technology