Nine languages, detected automatically and kept sticky.
Agents switch between English, Bangla and Banglish mid-sentence; travellers write in Arabic or Urdu. TravoPilot detects the language and keeps answering in it.
- The problem
- Most AI assistants handle English well and South Asian or Gulf languages poorly — especially romanised Bangla and mixed scripts.
- How it works
- Language is detected per message and held across turns. Banglish requests are answered in Bangla script; text is normalised so different spellings of the same Bangla letters match.
- Where AI helps
- The model understands intent in any of the nine languages, while every fixed string — wizards, hold prompts, booking cards — comes from localised templates, not improvised translation.
- Business value
- Serve customers in their own language without hiring per-language staff, and let agents work in the register they think in.
{
"service": "flight",
"trip": "OneWay",
"from": "DAC", "to": "DXB",
"date": "15 Dec",
"adults": 2, "children": 1,
"child_ages": // asked next — never guessed
}