Pharmacy Refill Copilot
An ML pipeline that interprets unstructured refill requests, classifies urgency, and surfaces risk signals — while keeping pharmacists in full decision authority.
problem
I spent years inside pharmacy operations at McKesson Specialty Health. First documenting clinical workflows, then leading digital product delivery. One pattern never stopped bothering me: the most dangerous moment in a pharmacy isn't when something goes wrong. It's when something urgent looks routine. A refill for insulin and a refill for vitamin D look identical in a queue. They are not identical in consequence. Staff spend 40–60% of refill processing time on manual triage. Urgent requests for anticoagulants or insulin don't surface faster than routine vitamins. A third of requests arrive with missing information. Faxed prescriptions are often illegible. And audit trails? Reconstructed after the fact, if at all. Rules-based automation fails here because clinical context varies wildly, edge cases represent 25–35% of volume, and the risk is asymmetric: missing an urgent request carries catastrophic downside, while a false alarm simply adds one item to the pharmacist's review queue.
solution
A multi-stage ML pipeline that processes requests from any channel — app, fax, phone, manual entry. The system extracts structured data via OCR and NER, classifies urgency into three tiers with transparent confidence scores, and recommends one of four actions: Process, Clarify, Escalate, or Hold. Every recommendation ships with human-readable rationale. Every AI output shows its reasoning. Every human override is captured and fed back into the model. The system never blocks operations: if the ML pipeline fails, workflows degrade gracefully to manual processing with zero downtime.
year
2025
timeframe
In progress, expected end of Spring 2026
tools
ML/NLP: XGBoost (urgency classification), BioClinicalBERT (clinical entity recognition) Data Processing: Python, Azure Document Intelligence (OCR) Design: Figma (pharmacist-facing interface) Project Management: Jira
category
Full-Stack