Mira
Live
Browser-based voice agent (Tavus CVI) backed by a FastAPI service that runs seven tool-calling booking functions, persists transcripts, and writes an LLM call summary.
The problem. When a voice agent can actually do things (book, retrieve, modify, cancel an appointment mid-conversation), the LLM’s tool calls must reach real services. They must also come back fast enough to keep the conversation natural.
What I built. The browser holds the call. Tavus CVI runs voice and a talking-head replica. When the model emits a tool_call, Tavus broadcasts it as a Daily app-message. The Next.js frontend dispatches it to a FastAPI backend that I wrote. The backend runs one of seven booking tools against Supabase Postgres and replies via conversation.respond. At the end of the call, the service writes a Gemini summary of the transcript.
The decision that mattered. I split tool execution across the frontend (dispatch) and backend (business logic). I did this to keep the round-trip inside the same Daily session instead of bolting on a separate signaling path. So the agent stays responsive while still talking to a real database.
Shipped, with the rough edge stated. The live frontend runs against the deployed backend. Tavus’s free tier caps conversation-minutes, so a recorded walkthrough covers the full happy path when a live session is rate-limited.