I poisoned 20 drug names to find out whether a local model belongs in a pharmacy
Description and use
The workflow reads handwritten prescriptions, interprets drug names with a local gemma3:4b model, applies deterministic safety checks, and flags uncertain lines for human review, improving accuracy and reducing manual effort in a pharmacy.
Problem
A neighbourhood medical shop in India needs a reliable way to read handwritten prescriptions and decide which lines can be filled automatically and which require human review, without relying on network connectivity.
Outcome
The system processes a prescription in ~58 s on an eight‑core laptop, flags all lines for review due to low confidence, and demonstrates that a local LLM can correctly interpret drug names when combined with deterministic safety checks.
Workflow
- Photograph handwritten prescription
- Run macOS Vision OCR to extract text
- Send extracted text to local gemma3:4b model via Ollama for drug name interpretation
- Apply deterministic gate checks (safety, grounding, vocabulary)
- Flag lines below confidence threshold or failing checks for human review
- Human reviews flagged lines and completes dispensing
Evidence
The whole pipeline runs from one command on the shop's machine: python3 scripts/run_page.py prescription.jpg
The OCR reads. The model understands. The gate decides. Only one of those three is allowed to be wrong quietly, and it is not the model.
- Record date: 2026-10-17T05:41:25.524Z