dzivkovi/video-intel

README preview

# Video Intel

> **30 seconds to read a mind map vs. 30 minutes to watch the video.**
> Scanned 15 videos from a single channel in under 2 minutes, ~$0.15-0.25 each.
> Free tier covers 8 hours of YouTube video per day.

Multimodal video intelligence powered by Gemini. Scan YouTube channels,
generate thematic mind maps, and produce enriched transcripts that capture
what was said AND what was shown on screen.

Then, optionally, it turns that whole corpus into a small analytics store and asks it questions a search box cannot - who reached an idea first, which creators genuinely cluster, what is suddenly spiking - with the statistics done honestly: real null models, significance testing, and a plain caveat on every number. See [The intelligence layer](#the-intelligence-layer-optional).

## Key Principles

- **Multimodal, not transcript-based.** Gemini sees video frames at 1 FPS,
  reads all on-screen text, and hears audio simultaneously. When a presenter
  says "as you can see here," the output tells you what was actually shown.
- **Decoupled task prompting.** Transcription (audio) and speaker identification
  (vision) run as separate tasks within a single prompt to preserve attention
  quality, borrowed from Laurent Picard's research.

View full repository on GitHub →