Explanation: How AI Scoring Works

ClipPulse AI ranks moments by virality potential using a two-stage pipeline.

Stage 1 — Transcription

The audio is transcribed locally with Faster-Whisper (a quantized Whisper implementation). This yields timestamped words in the chosen language. No audio leaves your machine.

Stage 2 — Moment selection (Gemini)

The transcript is sent to Gemini 3.6 Flash, which is prompted to:

1. Split the video into candidate segments (hooks, punchlines, insights).
2. Score each on a 0–100 viral potential scale.
3. Return start/end timestamps + a short reason.

The model is instructed to favor: strong openings, emotional peaks, clear payoffs, and self-contained context (so a clip makes sense on its own).

Stage 3 — Rendering

Each selected segment is:

• cut with FFmpeg,
• reframed to 9:16 (center-crop, with padding fallback),
• overlaid with animated ASS subtitles (word-by-word highlight),
• given a generated thumbnail.

Why local transcription + cloud scoring?

Keeping transcription local protects privacy and avoids per-minute API cost for the heavy lifting, while the (cheap) Gemini call does the creative ranking where a large model shines.

Tuning

You can change clip_count, the transcription language, and (in code) the Gemini prompt in backend/pipeline.py to bias toward different content types (gaming, education, sermons, etc.).