Cuff-less Blood Pressure (BP) monitoring has gained interest of the research community in recent years, due to its importance in continuous and non-invasive monitoring of BP for early detection of hypertension, thereby reducing mortality. Several approaches that involve photoplethysmography (PPG) and Pulse Transit Time (PTT) have been explored with promising results; however the requirement of two sensors makes them obtrusive for continuous use. Single PPG sensor approaches using machine learning have also been attempted, but there are certain deficiencies in these methods as they go for a one-size-fits-all approach. In this work, we develop an ensemble of BP prediction models based on demographic and physiological partitioning. Also, we incorporate a set of unique PPG features into our models, which results in test accuracies of 5 mmHg Mean Absolute Error (MAE) for Diastolic BP, and 6.9 mmHg MAE for Systolic BP. Given our marked improvement over ubiquitous models (18% for Diastolic BP and 11.5% for Systolic BP), this approach opens up avenues where single PPG sensor based methods can predict BP with a high degree of accuracy. This is a big step towards developing continuous BP monitoring systems, and can help in better management of cardiac health.