Conversational AI

Overview

This case records a product brief, not a verified launch. The intended product is a spoken dialogue in Russian on a phone, with a 3D avatar, a short response delay, and the ability to interrupt a reply. I will not describe those problems as solved until the status of the work is confirmed.

Context and challenge

Voice makes delay obvious. A pause that would pass in chat feels like a broken line. Russian adds its own pressure: recognition, phrasing, and a voice that does not sound like a translated assistant. An avatar makes the delay visible a second time, because a face that talks too late is worse than no face. Interruption matters just as much. If the person starts speaking and the reply keeps going, the product is not a conversation.

Reflections

The page stays short on purpose. Latency budgets, the realtime API, the interruption behavior, and the avatar pipeline are design problems I can describe. They are not results. The status of this work needs to be confirmed before any of that is written as something that shipped.