Nowadays machine learning algorithms allow us to solve various difficult problems and optimize things a lot. One of such problems is detection of facial expressions. Detection of facial expressions can be used for market analysis, self-driving cars and in the entertainment industry. However, many challenges arise when trying to build a classifier for custom emotion, such as covariate shift and imbalanced number of instances per class. In this paper we present a new image dataset OAHEGA, that consists of six emotions, and conduct experiments on it. We also provide a comprehensive study of building a successful application on top of a model trained using the highlighted dataset.