The Internet of Things envisions a web-connected infrastructure of sensing and actuation devices. However, the current state of the art presents another reality: monolithic end-to-end applications tightly coupled to a limited set of sensors and actuators. Growing such applications with new devices or behaviors, or extending the existing infrastructure with new applications, involves redesign and deployment. A proposed approach breaks these applications up into an equivalent set of functional units called context engines, whose I/O transformations are driven by general-purpose machine learning. This approach decreases computational redundancy and complexity with a minimal impact on accuracy. Researchers evaluated this approach's scalability--how the context engines' overhead grows as the input data and number of computational nodes increase. In a large-scale case study of residential smart-grid control, this approach provided better accuracy and scaling than the state-of-the-art single-stage approach.