Four captions from one clip: formal, sarcastic, humorous-tech, and humorous-non-tech. A vision model on the Fireworks AI API watches the frames so every caption stays accurate, then re-voices them, with a built-in LLM judge scoring accuracy and tone.