StyleCap: Grounded Four-Style Captioning

StyleCap is an accuracy-first video captioning agent for AMD Developer Hackathon ACT II Track 2. It turns unseen 30-second to 2-minute videos into four requested voices: formal, sarcastic, humorous-tech, and humorous-non-tech. Creative captions often drift away from what the video actually shows. StyleCap prevents that by separating observation from writing. Gemma 4 first converts chronologically sampled frames into a structured fact sheet containing the setting, visible entities, actions, and timeline beats. A second batched Gemma call produces three candidates for every requested style. A final multimodal selection call re-checks representative video frames directly and chooses the caption that best balances factual accuracy, concrete scene coverage, and unmistakable tone. The scoring path accepts arbitrary evaluator video URLs, validates the exact Track 2 JSON contract, supports every underscore style ID, and rejects missing or unexpected outputs before writing results. It runs inside a public Linux AMD64 container and uses a dedicated Fireworks deployment of Gemma 4 31B IT NVFP4. The repository includes deterministic mock inference, bounded downloads, cached frame extraction, automated contract tests, and a public Streamlit demonstration using the same production pipeline. The result is concise creative captioning in which humor changes the framing without inventing the scene.