After a shoot, the useful details end up scattered: camera cards, sound rolls, notes, reports, bins, exports. By the time the grade starts, someone is often doing the same search again - the take, the line, the setup, the look.
This is our attempt to keep that work attached to the media. The first version takes camera originals and sound files, makes proxies, syncs audio, transcribes dialogue and indexes the result. Slate OCR, focus notes, visual tags and report reconciliation come after the core run is dependable.
How it works.
A shoot moves through a chain of repeatable stages. Re-running a job only fills in the work that is missing.
Ingest, probe, proxy, sync and transcribe run end to end today, with face detection available per job. Stand-in backends keep local development and CI simple; production points the same code at the real render and transcription services.
Slate, visual tags, focus and reconcile tables are in the schema, but are not driven by a flow yet. They are the next layer, not part of the default run.
The point is not to make editors live in another tool. The metadata should come back into the edit and the grade as bin columns, markers and sidecars.
- Avid Media Composer - ALE, the standard dailies log: scene, shot and take straight into the bin.
- DaVinci Resolve - metadata CSV today; transcript, face and focus markers pushed to the render seat next.
- Premiere Pro - XMP sidecars, on the way.
Clips are keyed by source file, so the hand-off stays close to the way assistants already relink and rebuild bins.
How it's built.
A small orchestration service talks over HTTP to the heavier render and transcription services.
Transcription and proxy rendering run outside the core app. Rendering is routed per clip: FFmpeg for common codecs, Resolve for RAW formats FFmpeg cannot decode cleanly. The core asks for a connector by type; config picks the implementation.
Roadmap.
The order is deliberately boring: make the core run reliable, wire in the real backends, then add more metadata.
This is still in development, so the order can move. The dependency chain is what matters.
- Ingest, probe, proxy, sync and transcribe run as one chain. Faces can be requested per job.
- Screenplay parsing creates scene records for the shoot.
- Filename and transcript search work on stand-in backends, so local development and CI do not need a GPU.
- Wire in the FFmpeg renderer and Resolve render seat for production proxy generation, including RAW.
- Move to a Prefect server with per-backend work pools, so runs survive restarts and clips fan out in parallel.
- Add Alembic migrations and scheduled database backups.
- Add slate OCR, visual tags, focus timelines and a reconcile pass against department reports.
- Add semantic and face search on pgvector, with a Qdrant offload path if the embedding volume outgrows Postgres.
- Open the source under AGPL-3.0 once the tool is past v0.1.
Built at The Studio in Dunkeld by a director of photography who also works as a platform engineer. The tool comes from doing the dailies work, not from trying to invent a new category.
Licensed AGPL-3.0-or-later. The FFmpeg render path lands before the source opens, so a self-hoster can cover everyday codecs without a Resolve licence. Resolve stays optional for RAW. Source opens once it is past v0.1.