Small and Medium Enterprises (SMEs) constitute the backbone of the United States economy, accounting for approximately 44% of economic activity and nearly half of all private sector employment. Yet, these enterprises remain disproportionately exposed to financial volatility, operational disruptions, and fraud related losses, largely owing to constrained access to sophisticated analytical infrastructure. This paper presents a unified AI driven predictive analytics framework termed the FinResilience Architecture designed to simultaneously address financial forecasting accuracy, real time fraud detection, and operational resilience for US SMEs. The proposed architecture integrates Long Short-Term Memory (LSTM) networks and Transformer based temporal models for multivariate financial forecasting, Graph Neural Networks (GNNs) and ensemble anomaly detection algorithms for transaction level fraud identification, and federated learning with explainable AI (XAI) modules to ensure data privacy and decision transparency across operationally distributed SME environments. Validated through a mixed methods research design combining simulation, case study, and empirical survey data from 215 US SMEs across five industry verticals, the framework demonstrates substantial performance gains: a 31.4% improvement in 12 month revenue forecast accuracy over traditional ARIMA based baselines, a fraud detection F1 score of 0.947 on imbalanced transaction datasets using SMOTE augmented Graph Attention Networks, and a 28.7% reduction in mean operational recovery time following disruption events. The findings underscore AI powered predictive analytics as a transformative, scalable lever for SME financial sustainability and organizational resilience in an increasingly uncertain macroeconomic landscape.