I built Video Helper, an open-source tool that processes videos (YouTube, Bilibili, or local files) through an AI pipeline and outputs structured knowledge: mind maps, timestamped summaries, and practice questions.

The core pipeline: 1. Download/receive video 2. Extract audio with FFmpeg 3. Transcribe with Whisper 4. LLM analyses transcript → extracts concepts → builds mind map tree 5. LLM identifies semantic transition timestamps → FFmpeg captures keyframes 6. Frontend displays mind map + summaries + synced video player

The bi-directional navigation was the part I'm most satisfied with: clicking a mind map node scrolls to the matching summary block; clicking any summary block seeks the video player to that timestamp. This makes reviewing long videos much faster without re-watching.

There's also a mode where the backend runs headless as a "skill" inside AI editors (Claude Code, GitHub Copilot, etc.), where the editor's LLM handles reasoning while Video Helper handles transcription, storage, and visualization.

Tech: FastAPI + Python 3.12 + Whisper + SQLite / Next.js 16 + React 19 + ReactFlow + Tiptap / Electron desktop

Deploy: Docker compose, desktop installers (Win/Mac/Linux), or clone source.

GitHub: https://github.com/LDJ-creat/video-helper License: MIT