The system addresses the next-view recommendation problem: given the current view and conversational history, which views should be recommended to maximize exploration utility? Our solution combines three signals: data interestingness (𝐼 data ) computed entirely by the OLAP engine through deviation analysis, preference alignment (𝐼 pref ) estimated by the LLM from conversational context, and diversity (𝐼 div ) based on graph distance to previously visited views. This demonstration presents the system through three realworld scenarios spanning retail analytics, manufacturing quality control, and environmental monitoring. Conference attendees will interact with actual cubes containing thousands of records, observe how natural language queries influence recommendations in real time, adjust utility weights to see their immediate effect on suggestions, and explore navigation graphs that visualize their analytical journey through the cube. Navigating multidimensional data cubes through OLAP operations remains manual and tedious, often causing analysts to miss important patterns. We demonstrate NavLLM, an interactive web-based system that democratizes data exploration by combining large language model (LLM) preference modeling with data-driven interestingness measures. Unlike end-to-end LLM analytics that risk hallucination, NavLLM employs a hybrid architecture: the LLM handles preference scoring and explanation generation, while a conventional OLAP engine performs all numerical computations. Our demonstration showcases four realworld cubes covering retail, manufacturing, and environmental domains. Attendees will see how the system guides analysts through natural language interaction, real-time recommendations, and visual navigation graphs. Experiments show NavLLM achieves 71% higher cumulative interestingness than LLM-only approaches while maintaining fully verifiable results. 2