What the hospital needed
The front desk takes the same calls all day: which day a consultant sits, when OPD opens, where the emergency ward is, what a bed costs, where to collect reports. The hospital's information lived in a scanned FAQ, notices and the staff's heads. Callers who could not get through called again, or came in person to ask.
What was built
We OCR-processed the hospital's scanned FAQ and documents into a 129-chunk knowledge base with 57 contacts (departments, wards, counters), and put an AI receptionist on the line with a formal Nepali persona. Three rules were hard-coded: no medical advice, anything that sounds like an emergency is transferred immediately, and if a schedule is not in the documents the AI says so and offers a call-back. Requests for appointments and anything unresolved became tickets for the desk.
What the pilot showed
- Department schedules, services and directions were answered correctly from the documents; a department query returns the right chunk in about 30–40 ms.
- Scanned PDFs need real OCR work: a scanned FAQ with Devanagari tables had to be cleaned by hand before it was usable.
- Exact-match retrieval matters for names: a hospital directory needs a name index, not only semantic search, or 'Manipal' does not find 'Manipal Teaching Hospital'.
- Callers switch between Nepali, English and Hindi inside one call; the AI followed.
- The desk wanted a roster it could edit itself, so the schedule became a dashboard table rather than text in a document.
What we run alongside
The healthcare capabilities we built for other lines carry over: a national directory of 349 hospitals with district-aware 'nearest' search, 15 blood banks and 11 ambulance services with phone numbers, and a drug-information agent that looks up 10,077 medicine brands. Drug-information line case study →
If your hospital wants a full deployment with appointment booking into your hospital system, that is scoped as an integration project. Hospitals and clinics solution →