I built a 1960s-style home telephone system. No more shouting “dinner is ready!” or “mommy, come to my room” across two floors. We wanted an intercom system for a while, but we didn’t want robotic text-to-speech (TTS) announcements blaring out of smart speakers unprompted, and we discarded the otherwise tempting idea of analog two-way radios with a total lack of privacy.

Thanks to Gemini, I landed on GPO retro phones connected to Asterisk. Setting up a new home lab server, configuring Asterisk SIP trunks, implementing voice ID speaker recognition and bilingual speech input and output, and piping audio into my personal AI agent harness were all doable within a few days with AI help. I don’t think I would have had the patience and knowledge required to do any of the parts otherwise.

Both kids absolutely love it. We placed phones across key rooms and even installed one out in the garden play house. Who would have guessed that a Gen Alpha child and a Gen Beta child in 2026 would get so much joy out of picking up a tactile phone handset? They constantly make calls between rooms or ring “the AI lady” in English or German to play music on the living room speakers. I especially love it when they pick up the phone to call and tell us they’re ready for bedtime.

Now that our home AI agent setup is getting even more complex with different users, permissions, and locations to call from, I had to learn how to protect and improve the overall system quality with evals. Gemini suggested I use promptfoo. Test cases like “find the address of our building management company” or “play ‘Let’s Get Together’” with location set to “living room” were a great start. They helped me find a number of bugs in the agent codebase, prompts that needed to be clarified, and I realized that the scripts the agent uses needed clearer help screens and more helpful error messages. It’s a great setup, and I can add new scenarios whenever we find ways to improve the agent’s quality.