Caption Craft AI

CaptionCraft AI is an advanced, high-throughput video captioning agent developed for the AMD Hackathon (Track 2). It is designed to generate highly accurate, style-aligned captions for audioless video clips across four distinct personalities: Formal, Sarcastic, Humorous Tech, and Humorous Non-Tech. Core Architecture & Pipeline Bypassing traditional lossy intermediate text grounding pipelines, CaptionCraft AI processes raw video keyframes directly using the state-of-the-art Qwen 3.7+ Vision-Language Model via Fireworks AI. To maximize efficiency under the strict hackathon runtime limits, the agent calls the model concurrently using a python-based ThreadPoolExecutor and utilizes the reasoning_effort: "none" configuration parameter to achieve rapid inference times (averaging under 11 seconds per video clip). Dynamic Diversity Engine To prevent the common agent defect where models repeat identical sentence skeletons across unrelated clips, CaptionCraft AI incorporates a Delivery Angle Rotation mechanism. For creative styles, it dynamically assigns unique structures (such as git commit messages, incident postmortems, or narrator monologues) while maintaining a rolling n-gram memory bank to actively block vocabulary overlap.