CineCaption: Multi-Agent Video Captioning Agent

CineCaption is a video captioning agent built by Team Remetico on serverless Fireworks AI models, with MiniMax M3 as the primary vision model. It watches a video clip and produces captions in four distinct styles: formal, sarcastic, humorous_tech, and humorous_non_tech. The pipeline works as a team of agents. First, a motion-adaptive frame sampler analyzes the video and spends its frame budget where the action happens, so a quiet clip is covered evenly while a busy clip gets dense coverage of its key moments, and the first and last frames are always kept. The vision model then writes a detailed fact sheet of the clip: subjects, setting, action timeline, visible text and landmarks. Four style writers run in parallel, each drafting six candidate captions tuned to its style. Four judges score every candidate on accuracy and style fit while looking at the frames, and weak winners get one rewrite based on the judge's critique. Engineering follows the rules strictly. All twelve videos are processed in parallel and finish in about six minutes, every API request is capped at 25 seconds, structured JSON output is enforced server side, and a triple fallback ladder plus a global watchdog guarantee that a valid caption is always written for every style of every clip. Nothing is hardcoded: every caption is computed fresh from the video content at runtime. The image is a compact 245 MB linux/amd64 container that reads /input/tasks.json and writes /output/results.json, tested end to end on twelve varied clips covering nature, urban, animals, people, sports, food, weather and technology.