Your callers already speak three languages in one sentence.
Customers message on WhatsApp — typed or as a voice note — and the assistant collects the booking details in the language they actually used. Every completed request lands in one table, ready to act on.
hi kifak, badde kiné la ménisque, l'après-midi bi Hazmié
Summary: Wants physiotherapy for a meniscus injury, afternoon, at the Hazmieh branch.
- Service
- Physiotherapy
- Branch
- Hazmieh
- Time
- Afternoon
- Body part
- Knee — meniscus
- Languages
- Arabic · French · English
Example output on a typical Lebanese voice note. The caller's exact words are kept, never translated away — for physio they are the clinical record.
How it works
Native audio, not a transcript
One multimodal pass reads the raw voice note, so code-switched branch names, French medical words and numbers survive instead of being guessed from a broken transcript.
Code owns the state
The model proposes field values and reply text. Deterministic code owns every state transition, enum check and completion — so a request is never "done" because a model said so.
One table to work from
Requests, clients and appointments in one console. Triage a whole day's queue from a single screen, replay the original audio, and propose a slot back on WhatsApp.
Ready when the next message arrives.
Staff sign in to see everything the assistant has collected since their last shift.