Traditional credit limit management relies on static rules, often leading to suboptimal risk-return outcomes. This paper proposes an adaptive machine learning framework for personalized credit limit optimization using multi-source transactional data. Our approach integrates supervised learning (CatBoost/XG- Boost) for initial risk segmentation with reinforcement learning (Q-learning) to dynamically adjust limits based on evolving customer behavior. The model incorporates traditional credit metrics, real-time transactional patterns, and alternative data sources, processed through metaheuristic feature selection. Experimental validation on synthetic datasets demonstrates 18.6% higher profit- to-loss ratios compared to rule-based strategies while maintaining regulatory compliance. The framework addresses data privacy through federated learning and ensures interpretability via SHAP values. Results highlight reinforcement learning’s effectiveness in balancing adversarial objectives of revenue maximization and default risk minimization, particularly benefiting thin-file borrowers. This research enhances the personalization of credit management by offering a data driven alternative to expert driven systems, so as to be able to dynamically optimize limits to individual customer.