"Hospitality.ai" — Insurance-Aware Care Navigator

During a medical emergency, nobody has time to read forty pages of insurance clauses — yet those clauses decide what a family pays. The worst surprise in the Indian market is proportionate deduction: pick a room over the policy’s cap and every associated charge is scaled down, often costing far more than the room difference itself.

An AI that summarises a policy is easy to build and dangerous to trust. The product had to be decision support that can prove every statement, and that says plainly what it doesn’t know.

Three components, each doing only what it should. A policy agent extracts the document into a validated schema with messages.parse() and Zod; the document goes in with line numbers, and every claim comes back with a verbatim quote and a line range that is re-verified against the source and stamped exact, fuzzy or unverified — the UI renders all three differently.

The hospital and room matching engine is deterministic TypeScript: room fit, proportionate deduction, sub-limits, co-pay, deductibles, out-of-network haircuts and the sum insured, ranked by a fixed weighted formula whose full breakdown is shown under “Why this rank”. The model only narrates the result, streamed in behind a list that has already painted. A four-stage journey copilot (admission → investigation → procedure → recovery) surfaces policy-grounded guidance, like the 48-hour pre-authorisation window, while the patient can still act on it.