Balancing personalized and diverse content is a major challenge for recommender systems: One must provide the user with diverse, previously unseen content, while at the same time match offered content with user-specific interests. To achieve this balance, we at Taboola combine NLP, similarity metrics and supervised deep networks into a single recommendation pipeline. In this talk I will discuss these elements and describe how their combined pipeline is used to optimize user-engagement within a personalized “interest-space”.