Pathway Automation
LLM-driven generation of complete patient notification pathways — push, email and SMS — from program content.
Health programs need a steady drip of notifications — a push message on day three, an email in week two, an SMS nudge when someone falls behind. Writing them by hand for every program is slow, and consistency slips. This system generates the whole pathway from the program's own content.
The pipeline loads a program's weekly articles, videos and recipes, assembles a prompt with a set of existing notifications as style references, calls an LLM, and parses a structured response into the exact shape the production database expects — all three channels, correctly scheduled, ready for a human to review and approve.
It started as a Streamlit prototype reading from CSV exports to prove the concept end-to-end. I then ported it to the production TypeScript stack with live database access, few-shot prompting and video-transcript-augmented generation so the copy reflects what the participant will actually watch.
Highlights
- Generates push, email and SMS variants in one pass, matching an existing schema so output can be inserted directly.
- Few-shot prompting against real approved notifications keeps tone consistent with what clinicians already sign off on.
- Video-transcript augmentation grounds the copy in the actual program content rather than the title alone.
- Exports to JSON, SQL and CSV so reviewers can work in whichever tool they prefer.