An end-to-end AI pipeline I built that turns raw multi-camera footage into an editor-ready package — cut clips, transcript, subtitles, and timeline files for DaVinci and Premiere.

13 n8n workflows orchestrating a FastAPI backend, Cloudflare R2, and four AI models. The highlight is the cost design — a cheap-to-expensive funnel (OpenCV → Gemini → Claude Vision) that narrows ~7,200 frames to the ~10 best moments for about $1 per project.

I'm honest about the limits: per-frame vision scoring is at the edge of what models do well today, so this is a strong first pass that saves real hours, not a finished edit.

Stack: n8n, FastAPI, Directus, PostgreSQL, R2, Docker, ffmpeg, Gemini, Claude, AssemblyAI. Built end to end.