Hospital Chain Cut Prescription Processing Time 94% with Multilingual OCR + LLM Pipeline

The Challenge

Pharmacy fulfilment took 8-12 minutes per prescription because pharmacists manually typed handwritten doctor instructions into the hospital system. Errors averaged 2.1% — clinically significant. The pharmacy backlog created OPD discharge bottlenecks.

Our Solution

Iedeo built an OCR + LLM pipeline that extracts drug, dose, frequency, route and duration from handwritten prescriptions in English and Tamil. Confidence-scored output routes to a pharmacist review UI for cases below 92% confidence; everything above goes straight to the dispensing system with audit logs.

Architecture & Stack

• Mobile + tablet capture (pharmacy app)
• Hybrid OCR: PaddleOCR (handwriting) + Azure Form Recognizer (typed)
• GPT-4o for medical-context entity extraction with custom drug-name vocabulary (12K Indian-market drugs)
• Drug-drug interaction check against local formulary
• HIS integration (custom HIS — REST + HL7 FHIR)
• PHI redaction in logs, encryption-at-rest with AES-256
• Pharmacist review UI with side-by-side image + extracted fields

Timeline

3 weeks discovery → 8 weeks build → 6 weeks pilot in 2 hospitals → 4 weeks rollout

“We deferred this project for years because OCR alone could not handle Indian doctor handwriting. The LLM layer made it real.”