4 AI personas watch your live show and react in real-time. A fact-checker. A sound effects guy. A comedy writer. A cynical troll. Inspired by the Howard Stern Show. Streaming alongside the conversation as it happens.
The challenge, straight from @jason & @lonharris on This Week in Startups
Each persona mirrors a member of the Stern Show staff — the fact-checker, the sound guy, the comedy writer, the troll. Each runs on a different AI model. No single point of failure.
Monitors the conversation for factual claims and provides corrections or background data in real-time. Searches the web mid-show to verify statistics, dates, and attributions.
The "chaotic or negative cynical" commentator. Says what the audience is thinking but won't type. Dunks with internet-brain energy.
Supplies background context and sound effects. Communicates through precisely timed sound cues and the occasional razor-sharp one-liner.
Generates one-liners and jokes related to the current discussion. Setup-punchline structure, callback humor, the one who makes you spit out your coffee.
The video plays right inside the app. Pause it and the AI pauses too. Resume and everyone picks back up.
Embedded YouTube player on the left, 4 AI personas on the right. One screen, full experience. Synced play/pause controls.
Audio streams to Deepgram's Nova-3 model over WebSocket. Sub-300ms latency. Live transcript scrolls beneath the video.
Every 90 seconds of new dialogue, all 4 personas fire in parallel. Each gets rolling context + their own memory for continuity.
Responses stream token-by-token via SSE into a 2×2 persona grid. Dark theme. Auto-scroll. Typing indicators.
Two LLM providers. Swap any model. No single dependency. Open source, MIT licensed.
Jason Calacanis and Lon Harris put out a challenge on This Week in Startups: $5,000 + a guest spot to whoever builds a live AI sidebar with 4 personas watching the pod in real-time. Open source. Ship it.
So we're shipping it.