Recommender systems aim at helping users to find relevant information in an overloaded information space. Although there are well known methods (Content-based, Collaborative Filtering, Matrix Factorization) and libraries to implement, evaluate and extend recommenders (Apache Mahout, Graphlab, MyMediaLite, among others), the deployment of a real-time recommender from scratch which considers a combination of algorithms and various data sources (e.g., social, transactional, and location) remains unsolved. In this talk, we report on the challenges towards such a recommender systems in the context of online of offline marketplaces. In particular, we describe our solution in terms of the requirements, the data model and algorithms that allows modularity and extensibility, as well as the system architecture to facilitate the scaling of our approach to big data for online and offline marketplaces.