Researchers propose a framework for robust whole-hand and wrist kinematic tracking using the wearable WULPUS platform with an A-mode ultrasound probe. The system addresses the regression of 23 degrees of freedom directly on the device, overcoming limitations of prior non-wearable systems. A compact multi-output convolutional neural network containing 11,285 parameters is employed alongside an incremental training strategy to enhance generalization. This approach reduces mean absolute error by more than 17% compared to non-incremental methods. The model is deployed on the WULPUS nRF52832 microcontroller, achieving end-to-end tracking entirely on-device. Inference consumes only 0.73 mJ with a latency of 29.1 ms. The system supports full operation within 33 mW, enabling up to 36 hours of continuous use. This method also reduces wireless bandwidth requirements by 88% compared to raw data transmission.