The continuous development of hardware has allowed immersive technologies to be accessible for application in various daily tasks. It is currently possible to access immersive experiences through mobile devices, head-mounted displays, and other technologies. Extended reality (XR) is the term that encompasses a spectrum of various immersive and interactive technologies. Interactions are part of the hard core of immersive experiences and provide the possibility of interacting with the world around us. However, there are currently some SDKs intended for machine learning, such as ML-Kit. This SDK includes functionalities such as face and object detection and tracking, pose detection, and so on. However, there is a diverse set of functionalities, which are not yet grouped within some reusable structure (API or SDK). Many of these functionalities have proposed solutions using deep learning techniques. A characteristic of these techniques is that they tend to be complex in their space-time dimension. For this and other reasons, they are usually more difficult techniques to adapt for real-time use and other limiting characteristics of devices that reproduce immersive experiences. This work aims to study a viable architecture in which a set of ML methods that can be adapted to work in XR environments can be organized and grouped. These methods will have special characteristics since they must be adapted to provide real-time responses to users’ interactions with the environment. That is, they must adjust to the characteristics and limitations of immersive environments. Within the study, some of the main SDKs for immersive environments are analyzed and several ML methods are analyzed, related to the areas of computer vision and speech recognition, which should be included within the proposed architecture.