Data-driven building energy modeling is an emerging solution to facilitate the implementation of energy-flexible buildings. However, its black-box nature hinders interpretation, including with respect to human-building interaction. This drawback may bring risks to occupants’ satisfaction under aggressive demand-side interventions. A modeling framework that successfully integrates occupancy inference with data-driven energy prediction can help to address these challenges without raising cost or privacy concerns. In this paper, we propose OccuVAE, which incorporates domain knowledge on human-building interaction into the black box of data-driven energy prediction, simultaneously inferring occupancy states from whole-building energy data. Its multifaceted capabilities are enabled by its architecture, consisting of both a Conditional Variational Autoencoder (CVAE), as well as an interpretable system sub-metering disaggregation module. We test OccuVAE on a synthetic office building subject to stochastic occupancy schedules and system operation. We demonstrate OccuVAE outperforms existing baselines for occupancy level extraction solely based on clustering energy-metering data (average F1 scores above 0.7 vs. baselines around 0.5). It also shows robust energy prediction performance for different prediction horizons while providing insights into system sub-metering disaggregation. We also demonstrate that it can recover occupancy level profiles from real-world energy use data of an office building, and we highlight necessary future steps to further address real-world challenges. This prototype is a critical first step toward holistic predictive operation leveraging both energy and occupancy flexibility.