AgenticInspect: Risk-Aware Segmentation Refinement with Hybrid Auditing for Electronic Interconnection Inspection in Smart Manufacturing

Abstract

Industrial visual inspection requires not only accurate segmentation but also reliable and traceable decision making, since incorrect automatic refinements may lead to downstream manufacturing failures. Although segmentation foundation models achieve strong segmentation accuracy, they provide limited support for assessing whether refinement actions should be accepted, rejected, or escalated before deployment. To address this challenge, we propose AgenticInspect, a hybrid-audited agentic inspection framework that reformulates segmentation refinement as an action-level, risk-aware decision process. AgenticInspect treats an edge-deployable segmentation model as a replaceable skill. Based on the output of the segmentation skill, AgenticInspect introduces three task-specific decision modules, including uncertainty- and geometry-guided candidate generation, train-learned gating for high-risk refinement actions, and hybrid audit for risk-controlled mask selection. The hybrid audit layer integrates deterministic safety rules, train-learned safety gates, and selective large language model (LLM) review. The proposed framework provides action-level risk control and auditable decision making, with a traceable rationale for accepting, rejecting, or escalating candidate refinements. AgenticInspect was evaluated on a labeled connector-inspection dataset. The proposed refinement pipeline improved IoU from 0.9189303 to 0.9190180. The full hybrid policy reduced degraded refinements from 14 to 12 while maintaining nearly all performance gains (IoU = 0.9190131). Audit-component ablation attributes the observed reduction in IoU-degrading outputs to the train-learned risk veto, while LLM review serves as a selective, structured, and auditable assistive interface rather than an independently validated source of automatic improvement. These results position AgenticInspect as a risk-aware inspection framework that converts static segmentation outputs into auditable refinement decisions.