Personalizing large language model (LLM) responses in enterprise environments faces unique challenges: heterogeneous data sources, strict on-premise constraints, and unreliable LLM outputs. We present PersonaPlugin, a production-ready framework that transforms real-world telecom signals—call summaries, app usage, and location patterns—into structured personas for LLM personalization. Our system processes 1M+ behavioral records daily, entirely on-premise. The framework comprises three specialized engines, with the call extraction engine achieving >99% parsing success through cascaded fallback mechanisms. We conduct a comprehensive two-track evaluation: (1) general personalization across 800 scenarios, and (2) call-context personalization across 400 scenarios. Our LLM-as-a-Judge evaluation, validated against human judgments (κ=0.71), demonstrates that persona injection significantly improves personalization: +24.0% in P-Score and +26.7% in Adherence over baseline. Integrating call context with personas yields +11% additional improvement, validating the synergistic value of multi-source persona construction.

As subscription-based business models gain importance across industries like e-commerce, content streaming, and OTT platforms, the demand for recommendation systems tailored to these services has grown significantly. However, these systems frequently encounter challenges such as data sparsity, imbalanced user engagement, and product interdependencies, which traditional recommendation models struggle to fully address. Existing models often fail to account for the unique aspects of subscription service, leading to suboptimal recommendations. To address these shortcomings, this study proposes a new hybrid architecture that addresses both data sparsity and imbalance. By integrating recommendation and classification models, the proposed approach aims to improve both the accuracy and diversity of product recommendations. The system is built on a Recency, Frequency, and Monetary (RFM) feature store and employs correction techniques such as negative sampling and Gaussian rank scaling to better capture user behavior and product relationships. Additionally, combining Wide & Deep and XGBoost models ensures strong performance across different data environments. Both offline and online evaluations show that the proposed model significantly improves recommendation accuracy, click-through rates, and business revenue, proving its effectiveness for subscription-based services.