Who Is Right?!

A comic, live argument fact-check party demo. Microphone audio travels over an HTTPS WebSocket to the backend, which opens a Gemini Live session. Gemini transcribes, identifies checkable claims, and calls a backend tool. One Jev classifier-fast decision per claim then moves the gauge.

This is entertainment, not a factual authority. Audio is processed live only; the application does not record or store it.

Run

python -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
GOOGLE_API_KEY=... uvicorn app:app --port 8080

Production defaults: 180 seconds per session, three sessions per IP per UTC day, 80 global sessions, 240 Jev calls and 160 searches per UTC day, and five concurrent sessions. Gemini 3.8 Live standard audio is $0.005/input minute and $0.018/output minute; the 80-session cap limits maximum microphone time to 240 minutes ($1.20/day before short spoken responses). Override the corresponding environment variables to lower the caps.

MIT licensed.