The Challenge
YouTube holds an enormous amount of expert conversation — podcasts, interviews, lectures, debates — but almost none of it is addressable. Finding a single claim means re-watching a two-hour interview, and general-purpose chatbots will confidently describe things a video never said. Manual note-taking closes the gap for one person; it does not scale to a body of research.
Our Approach
- 1
Indexed English captions across a global catalog of channels, so transcripts became full-text searchable with channel, video and date filters.
- 2
Scoped every answer to the sources a user explicitly attaches rather than the open web, so the model can only draw on material already in the project.
- 3
Bound each claim to a YouTube timestamp, letting any statement be clicked through and checked at the exact moment it was said.
- 4
Separated long-running investigations into projects and one-off questions into chats, and exposed the same index through an API and MCP for programmatic use.
What Changed
- Answers that carry citations back to the moment they were spoken, instead of assertions taken on trust.
- Research that spans several channels at once rather than one video at a time.
- A queryable interface to the catalog for teams building on top of it.
