An AI rep that calls small businesses and hands them to a human
Alex calls a restaurant owner about their payment terminal, works out whether they are worth a rep's time, and then does the part most voice bots skip: it asks permission, puts the caller on hold with music, dials a real salesperson, whispers a summary of the conversation into their ear, and bridges the two of them together.
It replaced a stack built on Vapi, which took the per minute platform fee to zero. It runs on the client's own EC2 box under systemd rather than a third party voice platform, so nothing about it depends on a vendor staying cheap.
The model itself remembers nothing between calls. Every session starts blank. What looks like memory is a one paragraph summary the model writes about itself at the end of each call, stored in SQLite and fed back in at the start of the next one. That was deliberate. Feeding whole transcripts back would bloat every call, and keeping it in my own database means resetting someone's history is one line of SQL instead of a support ticket.
- Twilio
- OpenAI Realtime
- Pipecat
- FastAPI
- SQLite
- nginx
- EC2