OmniClipper and ClipVault
Long video in, captioned vertical clips out, with the indexing, checks and posting around it.
OmniClipper downloads a long video, transcribes it, has an LLM pick the strongest topics, renders vertical clips with burned-in captions and face-tracked framing, and schedules them to YouTube, TikTok and X. ClipVault indexes long video into about 66,000 searchable, LLM-described moments and composes Shorts by query. A status dashboard behind a Cloudflare Tunnel and single sign-on watches both. AI Content Factory is an earlier desktop take on the same idea, built around a local model.
Most of the engineering is in the failure cases. Uploads go up private and stay private for at least three minutes while the platform's rights and age checks run, because a test upload reported finished after 40 seconds and claims arrived later; a flagged video is deleted and replaced. A rule the model kept reasoning its way around became a regex gate in code. A disk-full outage traced back to a cleanup job that only ran when one user opened their clip list.
- About 1,200 tests in OmniClipper and 1,550 in ClipVault Oct 2026
- 492 clips (9.59 GB) moved to object storage after a disk-full outage; disk use fell from 100% to 79% Aug 2026
- OCR of image-based subtitles: about 49 minutes for a batch Whisper would need about 14 hours to transcribe Sep 2026
- Uploads stay private for at least 180 s before going public, so late rights claims are caught Sep 2026
- Four concurrent text-to-speech calls at 2 threads each reach 1.01x realtime; one call tops out at 0.52x Sep 2026
Built with Python, Flask, PostgreSQL, SQLite FTS5, ffmpeg, Whisper (Groq), DeepSeek, Ollama, Cloudflare R2, YouTube Data API