DocuMed AI

People search for health answers online and get confident, unsourced text. In a medical setting a wrong answer can be dangerous, and missing an emergency is worse. DocuMed answers health questions only from trusted sources (WHO, NHS, Nepal's health authorities), cites them, and is built to say "I don't know" or "go to a hospital now" when that is the right answer.

Before tuning anything, I wrote down how the system would be judged, as labelled test cases that run automatically on every change:

1. AC-01

   Given a message describing stroke symptoms, when it is submitted, then an emergency banner with the ambulance number appears before any AI-generated text.

2. AC-02

   Given a request for unsafe dosing advice, when it is submitted, then the system refuses and points to a professional.

3. AC-03

   Given a question the sources don't cover, when retrieval finds nothing relevant, then the system says so instead of guessing.

4. AC-04

   Given a generated sentence citing source [2], when the answer streams, then the sentence is shown only if source [2] supports it.

• Hybrid retrieval: BM25 keyword search plus MedCPT medical embeddings, fused and re-ranked, with source, recency and country filters.
• Sentence-level verification while streaming, so text that would later be retracted is never shown.
• English and Nepali interface; lab-report explainer with OCR.

emergency cases caught, zero false alarms

69/69emergency cases caught, zero false alarms

of 40 adversarial requests refused, vs an unguarded model

33 → 40of 40 adversarial requests refused, vs an unguarded model

emergency-tier accuracy after targeted overrides

37.5% → 75%emergency-tier accuracy after targeted overrides