A maternal and newborn health service in India rebuilt its AI triage tool so a machine can finally show its work.
Noora Health runs a WhatsApp line where nurses field more than 50,000 caregiver queries a month, and their hardest call is deciding which messages mean send help now. An earlier system used a large language model to flag emergencies and explain why, but the explanations were reasoning chains too long to audit at scale, and any prompt tweak meant re-running a full evaluation. The team split the job in two: an LLM now pulls out symptoms and patient context using a vocabulary written by clinicians, and a separate deterministic rule engine decides whether that combination counts as an emergency. Recall climbed from 0.565 to 0.810 and F1 from 0.606 to 0.702, with the rules doing most of the heavy lifting.
The real win is not the accuracy bump, it is the paper trail. Clinicians can now check each stage individually and see whether a message was mistranslated, symptoms were extracted wrong, or a rule was simply missing, then patch the rule engine without retesting the whole system. Since deployment, the system has triaged over 152,000 queries, flagged 18.7 percent as emergencies, and clinicians have added 48 new rules on their own.
This is a useful rebuttal to the industry habit of asking a single LLM to both decide and explain itself. Chain-of-thought reasoning reads well but doesn't hold up as an audit log once volume gets real. Splitting extraction from decision-making is a plainer, more boring architecture, and boring is exactly what you want when the wrong call means missing a maternal emergency.