Driven by the advancements in Artificial Intelligence (AI) and cutting-edge communication technologies, the interest in Digital Twins (DTs) has grown in recent years. Today, DTs are investigated in a variety of Internet of Everything (IoE) use cases, thanks to their capability to optimize and predict the behaviour of their physical counterpart. Several application-driven vertical architectures have been proposed for DTs, but a general-purpose modeling approach is lacking. Although the research community is promoting holistic DT representations, practical software deployments of complex physical entities will be likely achieved through multiple interconnected DTs that together provide composite advanced services. In this paper, we present a user-centric architecture to deploy DTs in a flexible, dynamic, and reusable fashion, according to the requirements of the user applications. In our vision, complex DTs can be composed by integrating digital replicas (DRs) of physical objects with an AI-driven Service Layer. DRs are implemented as representation models collecting real-time data from the physical counterpart and therefore enabling vertical intra-twin communications. In paralle, a Service Layer allows to share knowledge and services among DTs through horizontal inter-twin communications, i.e., between DTs in the virtual space, which can be exploited to anticipate users’ needs. A toy example is presented to showcase the conceived architecture.