This demonstration presents an ultra-low-power real-time hand gesture recognition system for edge applications. The proposed design utilizes a hybrid classifier based recognition core for static gesture recognition and an error-tolerant sequence analyzer for dynamic gesture decision. The recognition core comprises a shallow decision tree and an Edge-CNN which categorizes different gestures only based on the pixel data at the edge of the hand region. As a result, the on-chip memory and computation intensity are dramatically reduced. The proposed system can recognize 24 dynamic hand gestures with an average accuracy of 92.6% with 184μW power consumption at 25MHz