Build 03 · Private alpha
Drop a stream URL. Empyre AI watches it end-to-end, finds the moments, frames the speaker, burns the captions, writes the hook. You review and post.
Cuts shipped from streams by
The pipeline
Every stage is a tool you'd pick yourself — yt-dlp, whisper.cpp, ffmpeg, Claude, Gemini. We don't reinvent them. We orchestrate them so you don't open six terminal tabs to ship one clip.
yt-dlp pulls best-quality MP4 directly. No re-upload, no waiting on cloud transcoders.
Local whisper.cpp with word-level timestamps. Brand vocabulary biasing fixes "FaZe" not "phase."
Claude Opus 4.7 reads the whole transcript with prompt caching, returns ranked candidate spans + hook drafts.
ffmpeg trims to keyframes, MediaPipe locks the speaker (or pillarbox blur), karaoke captions burn in.
Gemini Nano Banana draws a clickbait still from the loudest frame, with hook overlay.
MP4 + meta JSON in your library. Review, drag to your scheduler, post.