The widespread adoption of smart buildings and digital twin technologies enables the continuous and real-time collection of extensive data on key parameters and performance of buildings. The analysis of these data through techniques of probabilistic machine learning provides confident predictive models of the building performances, which in turn allow enhanced methodologies of sustainable and resilient building design. Building seismic performance is crucial for earthquake resilience. Beyond the primary objectives of ensuring safety and preventing collapse, the seismic performance of a building can be assessed by examining its ability to maintain functionality and minimize repair costs. Recent studies have demonstrated that resilient buildings exert a lower environmental impact over their lifecycle, due to reduced repair/maintenance stages. It follows that resilience is a key requisite for sustainable design. In this paper, an integrated design procedure is introduced for optimizing structural seismic safety and energy efficiency of buildings under imprecise and limited information. More specifically, seismic excitation is modeled as a zero-mean stationary Gaussian random process; to take into account epistemic uncertainties, the parameters of the underlying Power Spectral Density function are described as interval variables. Data on energy consumption are measured through a network of sensors. Given the available information, energy forecasting is accomplished through Recurrent Neural Networks. The propagation of uncertainties is accomplished through the framework of Sustainable and Resilient Engineering using HAZUS, a methodology developed by FEMA to estimate potential losses from natural disasters. The two performances, i.e., seismic and energy costs, are then combined with Multi-Attribute Utility theory. A building example shows the attractive features of the proposed procedure.