In many industrial and service-oriented data systems, entity records are frequently duplicated across heterogeneous sources due to the absence of global identifiers, restrictions on the use of personally identifiable information, and inconsistent data standards across platforms. These challenges are particularly pronounced in real-world environments where identifying information is sparse, heterogeneous, and unevenly distributed, making traditional identifier-based joins and learning-based deduplication approaches difficult to deploy or maintain in practice. This paper presents a graph-based engineering framework for entity deduplication under realistic industrial constraints. The proposed approach models entity records as vertices and encodes all available identifying relationships—such as shared attributes, historical associations, and cross-system links—as edges in a heterogeneous graph. Rather than relying on attribute similarity, distance metrics, or probabilistic inference, duplicated records are resolved through deterministic graph connectivity. As a result, entity unification can be achieved without labeled data, supervised training, or predefined similarity thresholds. The framework is designed to leverage all available identifying information, including sparse and low-frequency signals that are insufficient for conventional relational joins. Its effectiveness is demonstrated through empirical validation on publicly available, large-scale industrial datasets characterized by missing global identifiers and heterogeneous schemas. The results show that the proposed graph-based approach enables reliable and interpretable entity unification while maintaining scalability in distributed processing environments. The framework is applicable to a wide range of industrial domains, including service platforms, enterprise data warehouses, and large-scale information systems, where accurate entity deduplication is a critical prerequisite for downstream analytics and decision support.