.png&w=256&q=75)
🤓 Latest Submissions
CineCaption: Multi-Agent Video Captioning Agent
CineCaption is a video captioning agent built by Team Remetico. 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 specialized agents. First, a motion-adaptive frame sampler analyzes the clip 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. A GPT-5.6 vision agent then studies the frames and, in a single pass, writes a verified fact sheet of the clip (subjects, setting, action timeline, visible text and landmarks, and the single most distinctive detail) along with the formal caption. It states only what it can actually see and never invents details. Three style agents then write the sarcastic, tech-humor, and everyday-humor captions, each grounded strictly in that verified fact sheet, so every joke stays pinned to what really happens on screen. All four styles follow one rubric: accuracy first, then style fit, with hard bans on invented numbers, breeds, brands, and jargon. Reliability is built in end to end. If any call fails, the agent automatically falls back to a full serverless Fireworks stack (MiniMax M3 vision, Qwen cross-check, Kimi humor writer, and a strike, pick and repair selector), and finally to safe templates, so it never returns an empty caption. Each request finishes in seconds, the full batch stays well within the time limit, and the agent always writes valid JSON output.
13 Jul 2026
