Traditional e-commerce recommendation systems focus on predicting items to buy, often overlooking the evolving intent of users during a session. This work introduces a hybrid intelligence framework that predicts user intent (browsing, comparing, buying, or abandoning) directly from clickstream data and leverages these predictions to generate adaptive UX behaviors. The framework integrates a dual-task sequential model (LSTM / Transformer) for intent classification and next-click prediction, with a generative layer (LLM-based) that translates model outputs into dynamic interface adjustments. Using open-source clickstream datasets, weakly supervised labeling, and temporal, behavioral, and sequential feature extraction, the system achieves approximately 85% accuracy in intent prediction and demonstrates improvements in engagement and decision efficiency in simulated UX scenarios. This approach bridges predictive and generative AI, offering a novel methodology for intent-aware adaptive interfaces in e-commerce.