Brocacho Caption Agent
Brocacho Caption Agent is a Track 2 video-captioning pipeline built for the AMD Developer Hackathon. The container reads the official /input/tasks.json file, downloads each video, extracts timeline frames with FFmpeg at one frame every three seconds, resizes them to control model cost, and sends the visual evidence to Kimi K2.6 through Fireworks AI. The first model pass creates factual observations about setting, subjects, actions, sequence, visible text, audio context, and uncertainties. The second pass turns those observations into the four required caption styles: formal, sarcastic, humorous_tech, and humorous_non_tech. The design focuses on accuracy first, then style, so captions stay grounded in the video instead of guessing. The final Docker image writes valid /output/results.json, supports hidden 30-second to 2-minute clips, and is published as a linux/amd64 image for automated judging.
Category tags:.png&w=96&q=75){kind=small}
.png&w=96&q=75){kind=small}
.png&w=96&q=75){kind=small}
For Innovators & Creators