Video pipeline
Long uploads become ranked clips and platform-ready files through a staged pipeline. Exact task names and retries live in the application code; this page is the conceptual map.
Stages
- Ingest — upload or URL import; media lands in object storage; a video row is created.
- Transcription — speech-to-text (Whisper-class) produces segments used everywhere downstream.
- Scoring — multi-signal highlight detection (LLM, audio, scene, mood, hooks, chat, trends, loudness, and optional vision signals). See the public Signals section on this site for the product framing.
- Director & style — auto-direction (zoom, reframe, pacing) and caption presets interpreted into render parameters.
- Render — FFmpeg with allow-listed arguments; plugins can hook pre/post render.
- Publish — scheduled or immediate posts to linked social accounts with idempotent retries.
- Analytics — per-platform metrics roll up for feedback loops.
Async execution
Heavy stages run on Celery workers. Progress streams to the UI via realtime channels.