Recommender systems are usually designed around a single domain and a single interaction model, yet practical applications often require recommendations across heterogeneous domains whose data differ sharply in structure and signal strength. This paper presents a multi-domain recommendation system implemented as a unified Streamlit application covering movies, hospitals, books, places, careers, and learning roadmaps. Rather than forcing one algorithm onto every domain, the system selects a method according to the available signal: truncated singular value decomposition (SVD) for personalized movie recommendation; TF-IDF with cosine similarity for content-based movie and book recommendation; deterministic disease-to-specialty matching for hospitals; location and category filtering for places; stream-to-exam rules for career guidance; and a prerequisite- and progress-aware rule engine for learning roadmaps. Optional live retrieval supplies movie, book, place, and learning resources, while SQLite persists roadmap progress and administrator-managed content. On the MovieLens ml-latest-small dataset, under a chronological leave-one-rating-per-user protocol, the SVD model attains an RMSE of 1.012 and an MAE of 0.780, against 1.117 and 0.918 for a global-mean baseline—error reductions of approximately 9.35% and 14.99%. The paper distinguishes the real MovieLens benchmark from the reproducibly generated demonstration datasets used for hospitals, books, and places. The result is a transparent, modular prototype showing that heterogeneous recommendation problems can be served by domain-appropriate techniques rather than a one-model-fits-all architecture.