Chatty Kazakh

A tutoring bot has to talk like a person, since nobody learns a language from canned replies. And you still have to watch the delays, the cost per request, the voice handling and the switch between languages.

We built the text and voice flows inside Telegram, one shared codebase, and the infrastructure around the models. The interface stays in the messenger people already use.

The real product question was how to make the conversation feel natural. We worked on whose turn it is, on hints, on network failures, and on the moments when the model answers something else entirely.

The technical side is assembled like a product someone will maintain, not a weekend script: the apps, the shared modules and the infrastructure sit in one repository.