Nice! I'm guessing you split the video in frames, analyze one frame every X frames with a VLLM (Qwen3-VL?), and then search in the descriptions of all the frames to find the object?