What they did
Tired of scam calls on their mobile, szert built CallSitter and now sends every incoming call to it. Calls divert to a Telnyx number where an OpenAI realtime voice model picks up. A separate classifier reads the transcript as it grows and labels the call as scam, sales spam or genuine, and a fixed call flow tells the voice agent what to do: keep a scammer talking, politely decline a sales call and ask to be removed from the list, or take a message from a real caller. A second model writes a running summary for the owner. Each call shows up as a Telegram message that updates live, and the owner can reply there to steer the agent mid-call, flip the call's label, hang up, or take the call over themselves. Daily spend, call count and call length are capped.
How it went
The poster says it works well and that all their calls now go to the bot. They published recordings: two police-impersonation scams they believe were real people, one of whom talked to the bot for close to an hour and called back after the bot cut the call at its time limit, a crypto robocall, a cold sales call and a genuine call from an electrician. In one recording the bot ends up accusing the scammer of being an AI. The poster gives no figures for total cost or how many calls it has handled.
Worth knowing
HN commenters argued that answering at all confirms your number is live and may bring more calls, and that most scam calls are auto-dialers with nobody on the line until you speak. The poster's reply was that they mostly do it for fun and to cost scammers time. The setup needs a carrier that lets you divert calls to a Telnyx number, and the README sets a $20 per 24 hours spend cap after which callers hear a busy tone.