A persistent shortcoming of contemporary recommendation engines is the inability to account for the contextual variability of individual users: a person who contributes highly discriminating ratings within one product category may behave entirely differently in another, yet most latent-factor methods process every interaction through a single shared embedding space. This paper introduces DSRS (Domain-Sensitive Recommendation System), a two-phase framework designed to resolve this inconsistency without relying on pre-labelled category annotations. In the first phase, spectral bi-clustering autonomously partitions the user-item rating matrix into cohesive interaction blocks, each block serving as a proxy for an implicit interest community. In the second phase, a restructured matrix factorization objective incorporates learnable domain-bias scalars alongside conventional user and item parameters, enabling simultaneous capture of global preference trends and domain-localized deviations. We present a fully functional implementation of DSRS as a Java-based (JSP/Servlet) social commerce web application deployed on Apache Tomcat~9, comprising 29 distinct JSP pages, a MySQL-compatible relational database with six core tables, a real-time Chart.js analytics dashboard, and a social friend-recommendation engine. Experiments on synthesized composite versions of MovieLens-1M (partitioned by genre) and the Amazon product review corpus show that DSRS reduces RMSE and MAE by 3--5\% relative to standard collaborative filtering baselines, with the margin widening markedly in high-sparsity regimes where domain structure supplies the only reliable inductive signal.