<![CDATA[Large-scale wholesale transaction systems face increasing risks from suspicious purchasing patterns, abnormal customer behavior, and operational inconsistencies, while conventional rule-based methods may fail to identify previously unseen patterns. This study develops a decision-level hybrid risk classification model that combines expert-derived business rules with post-hoc anomaly decisions from Isolation Forest and DBSCAN. The modules are orchestrated in Apache Airflow and executed through a scheduled daily DAG for batch transaction monitoring. The model was evaluated retrospectively using 12,476 wholesale transactions recorded over 12 months in a single company. The business rules and ground-truth labels were elicited from the same expert pool; therefore, the separation between model development and evaluation was implemented at the data level through independent training, validation, and test subsets. The Hybrid Rule + Isolation Forest configuration achieved the highest accuracy at 87.5%, compared with 74.3% for the authors' rule-based baseline. The resulting 13.1-percentage-point gain should be interpreted as an internal ablation result rather than a comparison with a state-of-the-art external model. For operational efficiency, the automated workflow processed a batch of 1,000 transactions in 42 seconds, compared with approximately 3 hours of manual processing. These findings suggest that combining interpretable business rules with anomaly detection at the decision level can improve risk classification while retaining operational transparency. However, because the evaluation used data from only one wholesale company and shared expert sources for rules and labels, validation across independent companies and expert groups is required before broader generalization.]]>
