Media Plane — LiveKit Mic Dev Client

Start a local LiveKit server (livekit-server --dev), then python -m src.entrypoints.token_server and python -m src.entrypoints.agent dev, then open this file directly in a browser (double-click it — no build step required, it loads the LiveKit JS SDK from a CDN).

Every real component runs in this test — nothing here is mocked or bypassed:

• VAD (on-device Silero, via livekit-agents' built-in VAD) — detects speech start/stop for turn-taking and barge-in
• STT (Sarvam Saaras, streaming) — continuous speech-to-text
• LangGraph — the conversation state machine (a different one per Call type below — see the stepper)
• Small LLM — phrasing only (naturalizes what the graph already decided to say; never decides what to say or which state to go to). Provider is set by LLM_PROVIDER in .env.
• TTS (Sarvam Bulbul, streaming) — text-to-speech

Only the transport differs from a real phone call — everything above this line is the exact same shared pipeline code either way.

Pick a Call type below before connecting — each one is a genuinely different LangGraph flow (not just different data within one flow), selected via an explicit agent dispatch this page's token request triggers (token_server.py), carrying whatever fields that flow needs as call metadata. Support Follow-up handles an inbound support ticket; Doctor Journey Onboarding is the post-delivery doctor-consultation call; Auto Booking Confirmation confirms attendance (or reschedules) for an already auto-booked slot.

Every flow now also mirrors Hindi/Hinglish if the caller speaks it, or switches immediately on an explicit request ("please speak in Hindi") — the language badge next to the mic meter shows what was last detected for the call.