Keptbits reads a long video, finds the moments that stand on their own, and returns finished vertical clips with captions burned in. One token, one clip.