Architectural tactics such as heartbeat, resource pooling, and scheduling provide solutions to satisfy reliability, security, performance, and other critical characteristics of a software system. Current design practices advocate rigorous up-front analysis of the system's quality concerns to identify tactics and where in the code they should be used. In this paper, we explore a bottom-up approach to recommend architectural tactics based on latent topics discovered in the source code of projects. We present a recommender system developed by building predictor models which capture relationships between topical concepts in source code and the use of specific architectural tactics in that code. Based on an extensive analysis of over 116,000 open source systems, we identify significant correlations between latent topics in source code and the usage of architectural tactics. We use this information to construct a predictor for generating tactic recommendations. Our approach is validated through a series of experiments which demonstrate the ability to generate package-level tactic recommendations. We provide further validation via two large-scale studies of Apache Hive and Hadoop to illustrate that our recommender system predicts tactics that are actually implemented by developers in later releases.