I built this for Track 2 of the AMD Developer Hackathon. The goal was an agent that watches a video clip and captions it in four completely different voices: formal, sarcastic, tech humor, and everyday humor. Here's how it actually works, step by step. The agent grabs 6 evenly spaced frames from the clip, not the whole video, just enough to understand the scene without wasting time or tokens. Those frames go to one vision model call (Minimax M3, running serverless on Fireworks AI) that writes a plain, factual description of what's happening. Then that single description gets reused across 4 separate quick text calls, one per style. So instead of asking the model to be funny and accurate at the same time, it separates what's true from how to say it, which kept the four captions from drifting apart or contradicting each other. I tested it against the 3 sample clips provided (a busy street, a kitten, an office worker), but the real evaluation uses about 12 hidden clips covering completely different subjects: nature, sports, food, weather. So nothing here is hardcoded to the examples. The whole thing runs as a Docker image, capped under 10GB, finishing all 4 styles per clip well inside the 10 minute limit. I built and tested it using AMD Developer Cloud, and the final image is pushed publicly to Docker Hub so it runs the same way for judges as it did for me.
Category tags: