Video Captioning Agent

Video Captioning Agent is our Track 2 submission for the AMD Developer Hackathon. The system takes a video URL as input and produces captions in the four required styles: formal, sarcastic, humorous_tech, and humorous_non_tech. The pipeline downloads the video, uses ffprobe to measure duration, then samples frames with ffmpeg at roughly one frame every five seconds, clamped between 8 and 20 frames. These frames are sent to Claude Haiku 4.5 as a vision request to produce a factual scene description. A second structured-output call rewrites that description into all requested caption styles in one JSON response. This design keeps every caption grounded in the real video content while matching the requested tone. Structured outputs guarantee valid JSON with exactly the requested style keys, reducing the risk of malformed output during judging. The Docker image is public, linux/amd64, and verified with the same /input/tasks.json to /output/results.json flow used by the Track 2 harness.