In this work, we address the problem of providing fast and on-line households appliance load detection in a nonintrusive way from aggregate electric energy consumption data. Enabling on-line load detection is a relevant research problem as it can unlock new grid services such as demand-side management and raises interactivity in energy awareness possibly leading to more green behaviours.To this purpose, we propose an On-line-NILM (Non-Intrusive Load Monitoring) machine learning algorithm combining two methodologies: i) Unsupervised event-based profiling and ii) Markov chain appliance load modelling. The event-based part performs event detection through contiguous and transient data segments, events clustering and matching. The resulting features are used to build household-specific appliance models from generic appliance models. Disaggregation is then performed on-line using an Additive Factorial Hidden Markov Model from the generated appliance model parameters. Our solution is implemented on the cloud and tested with public benchmark datasets. Accuracy results are presented and compared with literature solutions, showing that the proposed solution achieves on-line detection with comparable detection performance with respect to non on-line approaches.