Decord is a video loader built on top of hardware-accelerated decoders (FFMPEG, Nvidia Codecs, Intel Codecs) that provides efficient random access for deep learning pipelines. It includes bridges for deep learning frameworks, supports synchronized video and audio decoding, and is optimized for the random access patterns typical during neural network training.