Applied AI for fair mortgage lending: A fairness-constrained gradient harmonization framework

US financial institutions automated mortgage lending decision systems operate on datasets that encode decades of structural inequality along racial, gender, and socioeconomic dimensions. The machine learning models trained on these datasets risk perpetuating such disparities in ways that violate both the Equal Credit Opportunity Act and the fundamental principles of distributive justice. In this empirical study, we introduce the Fairness Constrained Gradient Harmonization (FCGH) framework, a neural architecture that jointly optimizes classification accuracy and differentiable demographic parity penalty through an adaptive gating mechanism that dynamically rebalances the two competing objectives during training, preventing either from dominating the gradient signal and thereby achieving a principled navigation of the accuracy-fairness Pareto frontier. Utilizing a four-phase experimental protocol, we compared FCGH against performance, fairness and demographic dimensions on HMDA records computed over both individual protected attributes and their intersectional combinations. The obtained results showed that FCGH achieved competitive classification performance while simultaneously reducing the demographic parity difference across race to 0.0059 and outperforming all baseline models on the fairness dimension without requiring post-hoc calibration or threshold adjustment. The three-dimensional visualization of the accuracy-fairness trade-off surface revealed that FCGH consistently occupies a favorable region of the Pareto frontier that no single baseline model reaches across all experimental conditions.