CostGuard: Policy-Governed Cost Anomaly Intelligence for Cloud DevOps Pipelines

Cost optimization in cloud-native CI/CD pipelines remains fundamentally reactive, with actionable financial signals typically emerging only after deployment decisions are executed. This work introduces CostGuard, a policy-governed cost anomaly intelligence system that transforms FinOps from retrospective reporting into real-time, enforceable governance within the CI/CD lifecycle. The proposed architecture integrates a dual-view anomaly detection paradigm:(1) an attention-enhanced residual BiLSTM for temporal stage-sequence modeling, and(2) a GATv2-based graph learning component for dependency-aware pipeline behavior.These complementary views are fused through an adaptive ensemble mechanism that dynamically selects optimal routing strategies based on calibration stability, ensuring robust performance across heterogeneous environments. A key innovation lies in bridging machine learning outputs with policy-as-code governance using Open Policy Agent (OPA). Model-derived risk signals are translated into deterministic and auditable runtime actions—ALLOW, WARN, AUTO-OPTIMISE, and BLOCK—enabling proactive cost control directly within pipeline execution rather than post-hoc analysis. The system is rigorously evaluated under a fixed ten-seed reproducibility protocol (42–132) across three domains: Synthetic (D0), TravisTorrent (L1), and BitBrains (L2). Experimental results demonstrate consistently high anomaly detection performance, achieving up to F1 = 0.9606 and PR-AUC = 0.9763, while maintaining low variance and full trial completion under constrained hardware (RTX 3050, 4GB VRAM). Beyond predictive accuracy, CostGuard emphasizes operational viability—including reproducible experimentation, memory-safe execution, audit-ready evidence persistence, and seamless integration into modern DevOps stacks (FastAPI, PostgreSQL, Streamlit). This work establishes a shift from passive cost observability to active, policy-driven financial governance in software delivery pipelines, providing a scalable and production-ready foundation for next-generation FinOps systems.