This tutorial explores the application of artificial intelligence in wearable devices and its unique inherent design challenges on both software and hardware. Through the use case of electromyography-based prosthesis control, real-time sensing, signal processing, and machine learning pipelines will be detailed. Overall system design and hardware integration will be discussed as potential solutions to most recent challenges in the state-of-the-art. Rehabilitation engineering seeks to provide technological tools for patients to recover as much of their abilities as possible following a major trauma or consequences of debilitating diseases. In the case of a limb amputation, a prosthesis aims to physically replace a missing extremity while trying to recover most of its functionalities. The degrees-of-freedom (DoF) required from the prosthesis vary from a limb to another, with the hand being especially complex. It is used in a wide variety of tasks, from daily routines to specific activities and involve multiple joints with different axes of motion. A multitude of muscles also allow for modulation of motion amplitude and strength. As such, not only is the hand a complex tool to replicate mechanically, but the design of an appropriate control interface also poses a major challenge. Machine learning provides a powerful tool to interface body and machines. In the case of amputated patients, where each user shows distinctive physical characteristics and resilient muscle abilities, machine learning algorithms provide adaptive solutions as they learn from user-generated signals. This tutorial will present how to structure a standard machine learning classification problem and how it can be applied to biomedical systems such as pattern recognition in electromyography (EMG) signals. Different sensor and feature extraction methods will be presented for classification algorithms such as Linear Discriminant Analysis (LDA), Support Vector Machine (SVM) and Neural Networks. Challenges and strategies will be discussed to improve the state-of-the-art on user-centered needs such as ease-of-use and intuitiveness. To address those concerns, an approach leveraging high-density EMG (HDEMG) and deep learning will be presented. This method aims to use convolution neural network (CNN) learned feature extraction capabilities to better fit unique users and enables transfer learning strategies to alleviate the dataset recording burden of individual users. A walkthrough of the complete sensing, preprocessing and inference pipeline will be given. Experiment protocols, tools and metrics will be presented to evaluate classification accuracy and real-time control reliability. Furthermore, real-time and user-centered wearable biomedical devices provide a unique set of challenges for applied artificial intelligence. Hardware constraints and limited availability of data for given individuals require creative approaches and paradigm shifts in algorithm design and hardware integration. State-of-the-art solutions and hardware platforms will be reviewed and discussed in regards to their potential of further advancing research in various fields of application and materializing results in the real-world.