AI Podcast Clip Cutter

AI Podcast Clip Cutter is a local-first, human-in-the-loop MVP for turning long-form podcasts into reviewable 9:16 clips. It demonstrates the complete product loop from deterministic candidate discovery through AI-assisted boundary review, human editing, and raw or subtitled export.

Portfolio MVP — v1.2.1

Screenshot showcase

Projects dashboard

New project

Review editor

Exports

What this project demonstrates

• local-first AI-assisted podcast clipping workflow;
• deterministic candidate discovery followed by semantic AI review;
• Gemini/LangGraph-assisted boundary refinement;
• human-in-the-loop clip review and manual editing;
• raw and subtitled vertical export generation;
• reproducible Docker Compose + Airflow setup.

Roadmap

See the Beyond MVP backlog for planned work outside the current Portfolio MVP.

Portfolio MVP complete — not a production SaaS. The product runs locally with Docker Compose and Airflow. Candidate discovery is deliberately neutral: replay-interest peaks create windows, but do not perform semantic scoring. Gemini through the LangGraph boundary-review path is the only automated semantic evaluation route; backend validation and the human editor remain authoritative.

Product evidence

Full third-party source media files are not distributed with this repository.

The validated MVP produces raw and subtitled 9:16 MP4 exports from authorized source material. Public portfolio screenshots focus on the application interface, review workflow, orchestration status, and export results rather than externally owned media.

Verified product flow

YouTube → Airflow → Faster-Whisper → Pyannote → deterministic merger → replay-interest candidates → Gemini/LangGraph boundary review → human edit → raw/subtitled render

Candidate windows contain no semantic summary or heuristic quality score. The editor builds excerpts from canonical transcript segments that overlap the current clip bounds. Time positions use HH:MM:SS.s; clip durations use MM:SS.s.

Architecture

flowchart LR
  U[React editor] --> N[Nginx]
  N --> A[FastAPI]
  A --> O[Airflow orchestration]
  O --> W[Pipeline workspace]
  W --> T[Faster-Whisper transcript]
  T --> D[Pyannote + deterministic merger]
  D --> C[Neutral candidate windows]
  C --> G[Gemini/LangGraph boundary review]
  G --> V[Validated review state]
  V --> S[(SQLite)]
  S --> U
  U --> F[FFmpeg raw/subtitled render]

Loading

The canonical Docker start is:

First copy the example environment file and fill in the required values and secrets:

Copy-Item .\orchestration\airflow\airflow.env.example .\orchestration\airflow\.env.airflow

docker compose --env-file .\orchestration\airflow\.env.airflow up --build --detach --wait
docker compose --env-file .\orchestration\airflow\.env.airflow ps

Services: React/Nginx at http://127.0.0.1:5173, FastAPI at http://127.0.0.1:8010, and Airflow at http://127.0.0.1:8080. Native local mode is retained as a development/compatibility path, not the official demo. See docs/DEMO.md and docs/ARCHITECTURE.md.

Validation and limitations

The repository has an offline Python suite, React/Vitest tests, lint/build checks, and Docker health checks. The verified 6 August 2026 validation for this release includes the Airflow scheduler import, Hugging Face model-info, Pipeline.from_pretrained(...).to(cpu), healthy containers, and endpoint smoke checks. No audio inference, full E2E, real Gemini call, or YouTube download is claimed by this release validation.

The project is CPU-oriented, local-first, single-user, and not hosted. It does not guarantee viral clips. Speaker labels are anonymous and are not linked to faces. Face tracking is experimental/best-effort and is not an MVP acceptance criterion. Subtitle styling is deterministic; it does not claim speaker-dependent colors. There is no subtitle text editor or Render All.

Users are responsible for processing only media they own or are authorized to use.

License

Code is MIT-licensed.