Inspiration
After a hospital stay, patients go home with many new medicines and paperwork they struggle to read. WHO has reported that only about half of patients on long-term therapies take them as prescribed. For older adults and non-English speakers the gap is worse, and mistakes can cause real harm. I wanted to build something that makes discharge instructions clear, safe, and understandable in the patient’s own language.
What it does
RxRosetta takes the medicine lines from a discharge paper and builds a plain-language daily plan in English, Tamil, or Hindi, with read-aloud audio. It checks for dangerous interactions and duplicate medicines, with an official DailyMed link on every warning. A teach-back quiz confirms the patient understood, and a daily check-in uses rule-based triage to say “stable”, “call your doctor”, or “get emergency care”. A live caregiver feed shows doses taken, quiz results, and alerts.
How I built it
I used Python and FastAPI with a vanilla JavaScript front end. The key design choice is that no medical claim comes from a language model. A deterministic engine matches medicines (handling Indian formats like 1-0-1, brand names, and spelling errors) against a curated knowledge base, and uncertain items are flagged for a pharmacist. I also built an evaluation harness: on 40 synthetic notes (227 medicines) it reached 94.3% medicine recall and 100% precision, with the misses shown openly.
Challenges I ran into
Making safety explainable. I chose auditable rules and source links over a black-box model, and I handled messy real-world formats and typos without guessing.
Accomplishments that I’m proud of
A complete patient-to-caregiver flow in three languages, with a measured evaluation and honest limits, built solo.
What I learned
In healthcare, trust comes from grounding and transparency about uncertainty, not from model size.
What’s next for RxRosetta
A pharmacist-reviewed knowledge base, live drug-label lookups, reliable photo transcription, validation on de-identified real documents, and SMS or WhatsApp reminders.
Built With
- fastapi
- html/css
- javascript
- python
- speech
- web
