FinGuard-Lite: Resource-Efficient Trustworthy AI for Financial Risk Decision-Making in Emerging Economies

Financial institutions in emerging economies increasingly explore machine learning for credit risk, fraud detection, and other decision-support applications, yet deployment conditions can differ substantially from those assumed by high-resource AI systems. Compute capacity may be limited, connectivity and infrastructure may vary, and expert human review may be scarce. These constraints motivate financial AI systems that allocate computational and human resources selectively rather than applying the same inference strategy to every case. We introduce FinGuard-Lite, a resource-aware selective prediction framework for trustworthy financial decision support under constrained deployment budgets. FinGuard-Lite combines a lightweight predictor, calibrated uncertainty estimation, risk-aware routing, and bounded human review. High-confidence cases remain on the inexpensive automated path, while uncertain or higher-risk cases can be escalated when additional resources are available. We evaluate the framework on the public UCI Default of Credit Card Clients dataset across five random seeds, measuring predictive utility, calibration, rare-event recall, inference latency, memory consumption, and decision loss under different review budgets. Our experiments additionally examine controlled distribution shift and explicitly report cases where simple routing does not outperform confidence-based alternatives. Rather than treating predictive accuracy as the sole deployment objective, FinGuard-Lite provides a framework for studying the trade-off among predictive quality, computational cost, and scarce expert attention. The results highlight both the potential and limitations of lightweight selective prediction for resource-constrained financial AI and motivate more realistic evaluation of trustworthy AI systems for emerging-economy settings.