A local-first video pipeline that does 80 percent of the edit before a human touches it.
Status: Running in production on real shoots. Code: private. This repo: the public overview.
- Ingests raw footage straight off the card (with automatic backup) and never uploads it anywhere: the whole pipeline runs on local hardware.
- Transcribes with faster-whisper large-v3 on GPU, so word-level timestamps drive the edit.
- Detects cuts and silence, builds edit decision lists, and hands off real editor drafts for the human taste pass.
- Finishes with an automated caption pass on the approved cuts.
- One shoot cycle produced 80 machine edits and 31 editor-ready drafts from a single day of footage.
The machine does volume; the human does taste. The handoff into a real editor as native drafts (not burned-in renders) is the piece that makes the last 20 percent fast.
Python, ffmpeg, faster-whisper (GPU), editor draft formats.
No upload time, no per-minute SaaS fees, no footage on third-party servers, and it works on a plane.
It is wired to specific brand assets and account pipelines. The architecture above is the shareable part.
Part of BDP Industries · Full story at brockpivec.com