Show HN: I built an AI-powered late-night call-in radio show from my RV
I live in an RV in the desert and I built a system that generates AI callers who phone into my
late-night talk show. Each caller has a unique voice, name, backstory, job, vehicle, and
opinions. They know the local weather, road conditions, and what's happening in the towns around
southern New Mexico. Some are recurring characters who call back with updates on their lives.
The stack:
- FastAPI backend running the show control panel
- OpenRouter for LLM (caller personalities, dialog, topics) — mostly Grok and MiniMax
- ElevenLabs / Inworld for TTS with 25+ distinct voices
- Caller personality system with memory — regulars remember past conversations
- Live phone integration via SignalWire so real people can call in too
- Post-production pipeline: stem recording, gap removal, voice compression, music ducking, EBU
R128 loudness normalization
- Self-hosted on Castopod, episodes served from BunnyCDN
The callers aren't scripted. The LLM generates their personality and topic, then we have a real
conversation. I respond as the host, the AI generates their replies in real time with TTS. The
result sounds like actual late-night radio — someone calls at 2 AM to argue about Pluto's
planetary status, another calls about their divorce, another has a conspiracy theory about
fusion energy.
Real callers can dial in live and get mixed in with the AI characters. Nobody knows who's real.
Listen: https://lukeattheroost.com
RSS: Spotify, Apple Podcasts, YouTube
Call in: 208-439-LUKE
The code is a solo project — happy to answer questions about the architecture.