Buildings consume approximately 40% of global primary energy. Almost a fifth of this consumption arises from faulty or poorly operated systems. Monitoring the state of such systems can provide beneficial diagnostic insight in favor of more energy-efficient buildings. Non intrusive load monitoring (NILM) is the process of breaking down the aggregate power consumption of a building into its individual constituents. NILM can be performed using smart meter data to provide more granular information on building appliances. It is crucial that the algorithms involved in this task be unsupervised since scaling such a process to millions of houses with human involvement is near impossible. It is also vital that the algorithms are compatible with the current, i.e., smart meters, infrastructure if they are to be implemented at a large scale and market-friendly.This thesis proposes a novel robust NILM algorithm capable of disaggregating the major loads for a portfolio of commercial and residential buildings from minutely recorded smart meter data. The key contribution of this thesis can be summarized in highlighting the major characteristics of power consumption in commercial and residential buildings, inspiring certain relaxing assumptions on smaller loads in favor of accurate state tracking for bigger ones. These assumptions address the seasonal and operational load variability in buildings and provide a framework for inference methods capable of producing good approximations for a computationally intractable problem. Finally, an empirical evaluation of the proposed NILM against several other studies in the field is performed. To this end, a unique dataset with labeled appliance-level data is created using data collected from several commercial buildings. The evaluation is presented for this dataset along with several other publicly available residential datasets. The results indicate an improvement in several evaluation metrics defined in this thesis.