# Keep Talking

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Park ChanJun](https://cerebralvalley.ai/u/cik666)
- **GitHub:** https://github.com/THEGREATCJPark/aipodcast-meta-hackathon
- **Demo video:** https://github.com/THEGREATCJPark/aipodcast-meta-hackathon/releases/download/keep-talking-v11/keep-talking-v11-antigravity-review.mp4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1527

Keep Talking is an AI research co-host built for small podcast teams that don't have a dedicated producer in the room. Good conversations lose momentum when hosts have to pause and Google something mid-recording. With Keep Talking, a host can toss out a quick fact-check question, keep the discussion moving, and get a concise, sourced answer only when there is a natural pause in the conversation.

In the demo, one host asks which Muse Spark model can be self-hosted. While Keep Talking looks into it in the background, the hosts keep talking about privacy, hardware specs, and local deployment. Once the lookup finishes, the AI checks where the conversation has moved before chiming in, delivers the answer, and hands the floor right back.

The core technical focus here is decoupling when a lookup finishes from when the AI speaks. Standard voice assistants treat every question as a prompt to talk immediately. Keep Talking tells the difference between casual banter among hosts and an actual request for information—and checks whether an answer is still relevant by the time it gets a chance to speak.

The prototype runs on React, Node.js, and LiveKit, handling per-speaker transcription, background web lookups, source tracking, turn-taking logic, and manual interruption controls. For this hackathon, I built the Meta-powered research and turn-taking pipeline on top of an existing recording studio codebase. The demo uses scripted synthetic host audio running against live Meta API calls, edited down for time.

Keep the conversation going. Let the AI handle the lookup.

## More from Meta Global AI Developer Hackathon

- [ProbeForge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1524)
- [Sentra-X](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1525)
- [Altovex CargoLinx](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1526)
- [Neural Referee](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1528)
- [OjO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1529)
- [MuZip](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1530)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1527. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt