AI video pipeline · personal tool
Unfold pipeline
A self-built pipeline that turns written scripts into short educational videos with AI-generated images and voice, renders them with Remotion and FFmpeg, and queues them for posting through n8n.
The problem
Turning a written script into a finished video with images, voice and captions, then posting it on five platforms, is slow manual work.
What it does
- A local API with 36 routes: render a video from storyboard JSON, produce voice and images, track jobs, manage a weekly content manifest and issue signed download links.
- Renders MP4 with Remotion and FFmpeg in two formats: 1080×1920 and 1920×1080.
- Scripts come from the Claude Code CLI, falling back to Meta’s Muse Spark and then to a local Ollama model.
- Images from Meta’s Muse Image; voice and word timing from a local, open-source VoiceStudio server.
- Publishing orchestrated by n8n workflows, with posting code for TikTok, Threads and Facebook and WhatsApp status reports.
How it is built
- Node.js and TypeScript with Fastify; Remotion for the video templates; MySQL with 17 tables.
- n8n for publishing, WireGuard for private access, nginx exposing only signed downloads and a health check, and macOS launchd.
- No Docker by design.
How AI is used
- Script writing with the Claude Code CLI, Meta Muse Spark or a local Ollama model.
- Illustrations with Meta Muse Image; speech and timing with a local open-source voice server.
Honest notes
- A personal, single-user tool built in about a week.
- Only the first parts of the first series have been published so far.
Engineering depth
Counts taken from the repository itself, not performance claims.
- 36
- API routes
- 17
- database tables
- 140
- storyboards
- ~36
- verification scripts