tinykitten_caption_generator

StyleCap watches a short video clip and writes a caption in four distinct registers — formal, sarcastic, humorous_tech, and humorous_non_tech — each faithful to what actually happens on screen. The hard part of Track 2 isn't describing a video; it's staying accurate while sounding unmistakably like each style, across wildly varied content (nature, urban, animals, people, sports, food, weather, technology). StyleCap is engineered to do both. How it works. A single-pass ffmpeg stage samples frames evenly across the entire clip (adaptive frame rate so a 2-minute video isn't truncated to its first minute) and extracts the audio track. Those frames are sent to a multimodal LLM that first grounds itself in the visible facts — subjects, actions, setting, on-screen text — then renders each requested style through explicit style contracts (do/don't rules, banned-word lists to keep tech humor out of everyday humor, and register exemplars). Every caption is constrained to only what's visible, so style never comes at the cost of accuracy. Built for reliability. The judged container is a thin, dependency-light async orchestrator — no model weights inside, starts in seconds. It seeds a valid, complete results.json at t=0 and rewrites it atomically after every clip, guards each clip with a timeout plus a global watchdog, and cascades through a tiered model fallback down to a template safety net. The result: it never crashes, never times out, never emits malformed output, and always returns a caption for every requested style — the failure modes that disqualified much of the field. It runs 12 clips end-to-end in ~36 seconds (well under the 10-minute limit), on a linux/amd64 image.