This paper introduces a novel architecture, the Great Individual TuTor Embodiment (GITTE) architecture, designed to generate highly personalized embodiments of Intelligent Learning Assistants (ILAs) through an iterative, participative approach. By leveraging Generative Artificial Intelligence (GenAI) models, the system dynamically adapts design features based on individual preferences, including those that are unconscious or explicitly rejected. The Pedagogical Agents-Level of Design (PALD) model serves as a structured taxonomy to collect and analyze design features at multiple levels, supporting systematic tracking of potential biases and hallucinations that may arise during image generation. To ensure compliance with the General Data Protection Regulation (GDPR), the GITTE architecture emphasizes local data processing, and pseudonymization, thus protecting student privacy while enabling real-time, dialogue-based exploration of embodiment preferences. The architecture’s modular design supports flexible integration of GenAI components, demonstrated by the seamless transition from one version of a Large Language Model to another without compromising functionality. A technical proof of concept validates the feasibility of creating ILAs that reflect the diverse design requirements of students, highlighting the system’s ability to account for age, gender, and other attributes. Iterative feedback loops further refine each embodiment, leading to continuously improved personalization. Nevertheless, ongoing challenges remain, including mitigating bias and improving scalability. Future research aims to expand the dataset for bias detection, optimize feedback processing mechanisms, and refine the PALD model to better collect nuanced design features, thereby improving the personalization and effectiveness of GenAI-driven ILAs. This approach holds promise for more inclusive, effective and adaptive digital learning experiences.