Context
ClipCAST is Jacques’s own podcast-clip system — not a client case study. Daily rotation pulls sources, renders face-aware clips, routes review through Telegram (and the ClipCAST Ops browser queue), then publishes via Postiz when accepted.
Approach
- Separate CLI pipeline (ingest / daily / review-poll) from a FastAPI read API and React review UI
- Keep humans in the loop: Accept / Reject with labels before publish
- Add scorer calibration so “viral” heuristics stay inspectable instead of mystical
- Run on Windows with Task Scheduler docs for production-ish daily cadence
Surfaces
| Surface | Purpose |
|---|---|
| clipcast.co.za | Public marketing site (podcast → Shorts positioning) |
| ClipCAST Ops dashboard | Product UI: review queue, ingest, episodes, published, calibration, settings (local demo against the FastAPI API) |
| Telegram review | Operator Accept / Reject before anything goes live |
| Pipeline stages | Discover → Score → Render → Review → Publish (Postiz → TikTok / Instagram Reels / YouTube Shorts) |
Outcome
Shipped as a working product codebase with dashboard and CLI, plus a live marketing site at clipcast.co.za. Portfolio hero uses a real ClipCAST Ops dashboard screenshot; pipeline SVG remains available as studio diagram media.
Stack notes
- Python package + CLI
- FastAPI backend
- React (Vite) ClipCAST Ops dashboard
- Telegram review drain
- Postiz publish path
- Whisper runtime documented for Windows