# editor.stumason.dev — the roll → crunch → edator pipeline

> Record once. Crunch makes it searchable. EdAtor cuts it. A self-hosted, agent-native
> video pipeline you own end to end. It works on my agent's machine.

Stu Mason (Folkestone) — a one-man production estate at platform-team scale. This is the
capture-to-cut pipeline: one screen-recording in, a finished branded cut — and a week of
shorts — out. Everything runs on hardware he owns. The pieces are open; copy the pattern.

## The chain

- **roll** — native macOS recorder. Captures screen + camera + mic on one shared clock,
  plus a telemetry track (clicks, keystrokes, scroll, cursor, and the Accessibility
  role/label of whatever was touched). Output is a deterministic "pack":
  `screen.mp4 · camera.mp4 · mic.m4a · metadata.jsonl · manifest.json`. The pack is the
  contract with everything downstream. https://github.com/StuMason/roll

- **Crunch** — self-hosted, OpenAI-compatible inference API (OCR, transcription,
  embeddings, rerank, caption). Reads the pack and turns the take into RAG-able,
  searchable data so edit decisions are cheap. No GPU, near-zero marginal cost, nothing
  leaves the box. https://github.com/StuMason/crunch · https://crunch.stumason.dev

- **EdAtor** — turns the searchable take into an EDL (the editorial decisions a human
  editor would make), renders to the frame with deterministic FFmpeg, dresses the cut in
  the private "Signal" overlay kit, and exports the finished video plus vertical shorts.
  AI does the judgement, FFmpeg does the work.

## Why

`click × on-screen text × transcript` = labelled action events. Once the footage is
searchable, an LLM can make real editorial decisions instead of guessing from raw pixels.
Own your stack, own your data, near-zero marginal cost.

## More

- /index.md — this page as markdown
- https://ai.stumason.dev — Stu Mason / the Signal design system