In this paper, we present the Nexus approach to efficient domain-specific integration of many loosely coupled data sources. A so called information maximizing mediation middleware (IMMM) has to cope with large data volumes and many queries, and at the same time achieve a tight semantic integration for the data instances. For efficiency and practicability reasons, we propose to use an extensible global schema and a limited domain-specific query language. This facilitates employing domain-specific semantic knowledge in the middleware: detect duplicates, merge multiple representations, aggregate and generalize information. Finally, we present a working prototype tailored to the domain of context-aware applications.