Generating natural, context-aware behaviors is crucial for effective Human-Robot Interaction, especially in contexts like Socially Assistive Robotics (SARs). Current methods often use complex architectures or preprogrammed behaviors, limiting adaptability. This paper introduces a novel, streamlined architecture for user data-driven context-aware behavior generation using a single Large Language Model (LLM), Gemini Flash 2.0. We show how one LLM instance can orchestrate both verbal and non-verbal robot behaviors, like adaptive speech and gestures, based on dynamic user characteristics. We evaluated this approach in a recipe suggestion scenario on the Pepper robotic platform, demonstrating its generalizability. While our architecture shows the potential of a single-LLM design for adaptive behaviors, a preliminary study on perceived social intelligence found no statistically significant differences between personalized and non-personalized interactions, indicating a need for further research. Our work contributes to more trustworthy and transparent SARs by exploring user data integration for context-aware behavior generation and identifying future research directions to enhance personalized and engaging interactions.