As online marketplaces continue to expand, recommendation systems have become essential tools for helping users navigate large inventories and discover relevant products. By reducing information overload and enhancing personalization, these systems not only improve the user experience but also support platform-level goals such as retention, engagement, and revenue growth. This thesis focuses on enhancing the recommendation engine for Encore, a video-based, auction-style second-hand marketplace targeting younger users. I evaluate and adapt a range of established techniques—including popularity-based baselines, pattern mining, content-based filtering, graph-based embeddings, and collaborative filtering—to identify the most effective strategies for Encore’s sparse and dynamic environment. In response to challenges such as limited user history and inconsistent metadata, I propose a recommendation system based on matrix factorization with implicit feedback, enriched by behavioral signals and semantic tags generated via Large Language Models (LLMs). Implemented in PySpark, the system is scalable and modular, allowing for future integration of new data sources and model components. Empirical results demonstrate that the proposed approach achieves a strong balance between personalization, novelty, and system scalability—key factors for supporting growth and user satisfaction on emerging digital platforms.