AI Content Pipeline
Raw video in, finished content suite out.
Hand it a raw, mistake-filled recording and it removes the retakes, adds overlays and music, and produces a publish-ready video along with the full content suite: YouTube description, companion blog post, social clips, and subtitles.
- 18 pipeline steps
- 2–4 hr of manual editing, before
- ~15 min per video, after
How it works
- Finds the mistakes ElevenLabs transcribes with word-level timestamps, the pipeline detects retakes and false starts, and MLT with FFmpeg cuts frame-accurately with no audio drift.
- A state machine with recipes The 18-step pipeline resolves its own dependencies, and recipes run just the part you need: subtitles only, music only, or overlays only.
- Each model where it's strongest Gemini handles large-context transcript analysis, Claude writes content and designs overlays, and GPT evaluates the results and serves as a fallback.
- Blog posts that cite prior writing A hybrid ChromaDB and BM25 index over an Obsidian vault lets generated posts reference earlier writing, and it classifies notes as original ideas, referenced material, or personal content.
- From a CLI to a platform A dashboard and worker queue: workers poll for jobs, process video, upload artifacts to Cloudflare R2, and report status, so it scales horizontally.