First we present Unison, a conceptual music recommender system for groups of people; the system aims at generating a playlist that takes musical tastes of all the group members into account. We discuss both theoretical and practical concerns related to such a system. We develop a model of user preferences and discuss how we can shift from individual recommendations to group consensus. In constructing the user preferences model we use an intermediary music track model that combines user-generated tags with a dimensionality reduction technique to build a compact spatial embedding of tracks. Secondly we introduce GroupStreamer, a practical implementation of the system that runs on Android devices. We present the technological choices that were made along the way.