Can AI replace diclip?

Build me a personal substitute for diclip, not a platform clone.

- Use Python 3.12 + FastAPI for a localhost web app, a plain JavaScript
  frontend, faster-whisper for word-level transcription, and FFmpeg for
  rendering.
- I import an MP4, MOV, WebM, or MP3 file and get a transcript synced to the
  video with speaker timing.
- A ranking step selects the strongest 30-90 second moments: call an LLM API
  key from .env with the transcript, or fall back to a local heuristic
  (keyword density, question marks, pauses) when no key is set.
- Show each candidate clip with its hook line, a confidence score, and a one
  sentence reason, marked ready or pending.
- Let me accept or reject each clip, edit in/out points, and choose 9:16, 1:1,
  or 16:9 output.
- Render clips with FFmpeg using center-crop or blur-pad, burning captions
  styled by speaker. Never overwrite the source. Show progress and a useful
  failure message.
- Store projects as JSON under ~/diclipDIY/projects and renders under
  ~/diclipDIY/exports, with a recent-projects page and a delete action.
- Bind to localhost only. No accounts, telemetry, or network calls after the
  model and transcript are local, except the optional LLM ranking call.
- Deliberately exclude face-aware reframing with seat tracking, the full
  in-browser timeline editor, cloud container rendering, yt-dlp downloads, and
  credit billing.
- Add unit tests for the ranking fallback and caption grouping, plus one smoke
  test that imports a short fixture and produces a playable MP4.
- Include a README with setup, data paths, and an honest note that CPU
  transcription and rendering are slow on long videos.