03 / Voice AI
- 1,000+
- Outbound calls a day
- 95%
- Intent recognition accuracy
The problem
Real-estate lead qualification ran on human telesales: slow to reach new leads, impossible to scale, and with no systematic way to learn which conversations converted.
The approach
- 01
Bidirectional WebSocket streaming on both the telephony and client sides, for low-latency speech, live transcription and real-time intent recognition.
- 02
A multi-provider voice stack: Whisper and Deepgram for speech-to-text, ElevenLabs and Google Cloud for text-to-speech.
- 03
A LangChain conversation engine built on state machines, so complex multi-turn dialogues keep their context across turns.
- 04
A post-call analysis pipeline that reads transcripts, identifies the prompts that are failing, and surfaces improvements, so call quality rises without anyone reviewing calls by hand.
- 05
A Next.js admin dashboard for monitoring live calls, with transcripts, keyword highlighting and audio sync.
Stack
Echo's architecture was shared work with my line manager.