Skip to content
All work

Authorized Clipper

An agentic clipping pipeline that publishes nothing on its own.

Media · Agentic pipelineClosed source

An agentic, human-approved video-clipping pipeline. A FastAPI media worker downloads, transcribes, scores candidate moments with an LLM editorial pass, and renders vertical cuts, feeding a Next.js dashboard where a person reviews and approves them.

Problem
Find and render strong vertical clips from longer source video.
Constraint
Automated editorial scoring has no authority to publish under a person's name.
Decision
End the automated pipeline at a human review-and-approval dashboard rather than at a publishing integration.
Evidence
Four automated stages—download, transcription, editorial scoring, and rendering—terminate at review, with no autonomous publish step.
FastAPINext.jsWhisperFFmpeg

What it is

A FastAPI media worker runs the automated half: download, transcribe, score the transcript for editorial moments with an LLM, and render vertical cuts. A Next.js review and approval dashboard is the other half, and it is not optional — the pipeline ends there rather than at a publish step.

The pipeline

  1. Download
    source media
  2. Transcribe
    Whisper
  3. LLM editorial scoring
    candidate moments
  4. Vertical render
    FFmpeg
  5. Review & approval
    Next.js dashboard — a person decides
Four automated stages, then a person. The dashboard is the terminal node by design: nothing downstream of approval happens without it.

Why it stops there

This follows the same rule as the payroll engine's posting step and the trading system's risk manager: a human stays in the loop for anything that moves money or publishes on someone's behalf. An editorial model is good at proposing a cut and has no standing to decide that it should go out under someone's name.