: Structure-based chromatographic retention prediction is critical for efficient method development and impurity characterization or monitoring in pharmaceutical analysis yet remains challenging without proper databases. Here, we present a practical project-specific quantitative structure–retention relationship (QSRR) framework that integrates descriptor fusion with a rigorously designed feature- and model-selection strategy to enable accurate retention time (RT) prediction using unfiltered miniature reversed-phase liquid chromatography (RPLC) datasets. Six molecular descriptor libraries, along with their various combinations, were evaluated using multiple machine learning algorithms within a nested cross-validation (CV) framework optimized to balance model fit and generalizability. The workflow was benchmarked using three late-stage Pfizer small-molecule projects, each comprising 17–21 compounds, including an externally validated set. High prediction accuracy was achieved across all case studies (R 2 > 0.96, mean absolute percentage error < 9%), with absolute RT errors below 2.5 minutes (for analysis time ranging between 30 and 60 min) for all compounds in the EV set, including regioisomeric impurities. Notably, unlike the conventional localized QSRR modeling, the project-specific QSRR modeling does not require explicit structural similarity filtering because the compounds for model development and external validation are inherently chemically related.