Talking Tutor

The relay is the whole design. The browser never talks to Gemini directly; it opens one WebSocket to the backend, which opens one Live API session with a system prompt rendered from the study guide. Automatic activity detection is turned off so a press of the talk button maps exactly to activity_start and activity_end, which keeps the model from answering background noise. The model is given two tools: record_answer, which carries the question, the student's answer, a correctness flag, and one sentence of feedback, and end_quiz, which carries a summary. The relay converts those into answer_recorded and quiz_summary frames the client renders as a live scoreboard, and writes the final score into quiz history. Pacing is enforced server-side: the prompt tells the tutor to end its turn after feedback, and the relay waits 2.5 seconds before prompting it to ask the next question, cancelling that prompt if the student starts talking. Cost control is layered: an origin check, a per-device and global daily session cap, a concurrency cap, a 3-minute clock, an inbound audio byte quota, and a 45-second idle timeout. When the clock runs out the relay locks input and gives the tutor up to 30 seconds to finish its turn before ending the session, and the client drains queued audio rather than stopping playback. Everything runs on a single Cloud Run instance so the caps can live in memory.