Personalized experiences in multimodal assistants rely on accurate user understanding, yet large-scale training for personalization remains limited by privacy constraints and data sparsity. We introduce a framework for generating Comprehensive Synthetic Personas (CSPs) and personalized synthetic training data through taxonomy-guided knowledge enrichment, in-context learning, and Chain-of-Thought (CoT) knowledge distillation for personal knowledge inference. This dataset is used to fine-tune a large language model (LLM) that learns to infer user interests and attributes from interaction histories. Evaluation shows that models trained on synthetic data outperform those trained on real de-identified data in precision and recall when evaluated on the same real de-identified test set, using LLM-as-a-Judge (LLMaaJ) and human annotation. This approach enables scalable and privacy-safe personalization learning, supporting downstream applications in AI assistants.