Glean

Ever save a video about a product, place, or idea to “watch later,” only to never revisit it? Or find a great recipe and end up scrubbing through the video just to write down ingredients and steps?

Glean turns videos into structured data stored directly in tools you already use (currently Notion). Users connect their databases, save videos via a browser extension, and Glean automatically extracts what matters from each video into the right schema—no manual replaying or note-taking required.

How we built it

We designed the data model and user flows in Figma, then used Stitch on Gemini 3 to generate a high-fidelity design. The frontend is built with React + Vite, while Supabase handles authentication, job management, and metadata storage. Our backend uses Python, FastAPI, and asyncio, with Composio managing third-party integrations like Notion.

Gemini 3 is central to the application and powers a multi-agent pipeline:

• Question Generation Agent: Uses Gemini 3 reasoning to generate onboarding questions for a given database schema.

• Prompt Generation Agent: Converts schema + answers into a frozen extraction prompt.

• Content Extraction Agent: Leverages Gemini 3’s video understanding, long-context reasoning, and function calling (Google Search and Maps) to extract structured data from videos.

• Critique Agent: Reviews outputs against the schema and provides corrective feedback.

What we learned

Composio significantly simplified third-party integrations, and with the right agent setup and tooling, building a full-stack app plus extension is surprisingly fast.

Looking forward

We’d love to support more data platforms, video sources (TikTok, Reels), and longer-form videos beyond YouTube Shorts. Additionally, if we have more time, we'd love to implement a feature to let users replay the processing of videos that got an error and edit the frozen prompts.

Built With

• composio
• gemini
• javascript
• python
• react
• supabase
• vite