Abstract

We present a novel multi-agent framework that integrates interpretable machine learning (ML) models with large language models (LLMs) to deliver personalized credit risk recommendations. Using the Lending Club Loan dataset, we train an XGBoost based credit risk model, applying SHAP (SHapley Additive exPlanations), we extract the top risk factors for individual borrowers. Subsequently, these personalized explanations are provided to an LLM, which generates actionable recommendations to help customers improve their credit profiles. A structured multi-agent LLM review pipeline is introduced to assess each recommendation for tone, compliance, and legal soundness. Recommendations scoring above a set threshold in all review dimensions are only shared with customers, ensuring fairness, transparency, and regulatory adherence. The experimental results demonstrate the effectiveness of the framework in providing clear, actionable, and compliant recommendations.