The energy system is rapidly changing in the hopes of reducing greenhouse gas emissions, installing large amounts of renewable energy generation and electrifying swaths of fossil-fuel based energy use with technologies like heat pumps and electric vehicles. These new loads combined with new variable renewable energy generation create challenging problems for regulating the power grid. This dissertation presents a scalable, data-driven control and modeling framework for making residential buildings grid-interactive, allowing them to act as distributed energy resources and provide additional controllability for grid operators. First, a new analysis of real-world, smart thermostat data motivates the problem by showing that uncoordinated, energy management tools can come with unintended, negative consequences like load synchronization. To solve this, three new methods are then presented for coordinating these energy management tools in a scalable way based on real data. Each focuses on solving three main challenges that have so far prevented smart building control from more widespread adoption: (1) building modeling difficulties, (2) control computational constraints, and (3) multi-agent coordination. These methods provide an end-to-end solution for providing scalable, data-driven control for grid-interactive buildings, reducing the barriers for adopting smart building control and mitigating the challenges that new electric loads will have on the grid.