Building production voice agents that don't embarrass you

A voice agent that handles negotiations with real humans is not the same as a chatbot. Here's the architecture that keeps Chasyr production-ready.

The state machine

Every call follows a finite state machine. Greeting → identity check → reason disclosure → negotiation → outcome → compliance close. The LLM generates natural language; the state machine gates which states are reachable. This single constraint eliminates 80% of failure modes.

Audio pipeline

Real-time audio requires careful buffering. We use Deepgram for transcription, pipe to Claude for reasoning, and use Elevenlabs for voice synthesis. Latency matters more than fidelity.

Fallback design

When the LLM gets confused, we escalate. No recovery attempts, no retry loops. A warm transfer to a human is cheaper than a bad negotiation.