The design problem
The interesting design problem, and the one that turned into a research paper, was what to do when the model’s prediction sits right on the boundary between risk levels: a doctor can read a 0.487 vs 0.499 split and shrug, but a mother handed a red alert can’t. The system solves this by making uncertainty part of the UI itself: derating an ambiguous “red” down to “amber” specifically on the mother’s screen, never hiding real risk, just refusing to overstate a coin flip as certainty.
Where it actually stands
Built with React/Vite on the frontend and FastAPI + Supabase on the backend, with a rule-based fallback so the whole thing runs standalone with no ML artifacts required. Two of three interfaces, ASHA and doctor, are fully wired end-to-end; the mother-facing view remains a UI mock, documented as such. That’s a deliberate, honest scope cut, not a hidden gap.
The hardest design question this prototype raised, how to communicate model uncertainty to a non-clinical audience without either alarming or misleading them, became a separate research paper, Three Voices, built as its own rigorous exercise on a public dataset rather than this system’s real data.