Up to 36x Speedup: Mask-based Parallel Inference Paradigm for Key Information Extraction in MLLMs | Signal Canvas | ScienceToStartup