Fiber-reinforced polymer (FRP) composites bonded externally to reinforced concrete beams have shown promise for increasing shear load-carrying capacity. However, accurately predicting the shear strength contribution of FRP remains challenging due to the complexity of the shear failure mechanisms and interactions between the contributing components. Existing design provisions have limitations in modeling this behavior. Recently, machine learning has shown promise for addressing such complex problems but lacks transparency. Therefore, this study developed a data-driven Constrained Monotonic Neural Network (CMNN) approach implemented using Python libraries Keras and Tensorflow, which integrates domain knowledge to provide reliable FRP shear-strength predictions. This framework directly encodes expected monotonic input-output relationships through constraints, harnessing both predictive power and monotonicity guarantees. A comprehensive database of 273 beam tests was used to train and evaluate the model. Bayesian optimization through the Optuna framework and 4-fold cross-validation were used to tune the hyperparameters of the CMNN model. Extensive evaluation demonstrated the CMNN model's superior accuracy, yielding an R² value over 0.9, a CI above 0.93 and a low MAE under 16 kN on both training and unseen test data. The model also outperformed existing design code provisions, achieving significantly optimal performance metrics across various strengthening configurations. The analysis of the predictions revealed logical sensitivity trends aligned with the principles of shear mechanics. Such physics-informed machine learning presents a next-generation technique for performance-based structural design, contributing an effective and interpretable tool to advance knowledge on quantifying the shear capacity contribution of externally bonded FRP reinforcement in concrete beams. • Develop a Constrained Monotonic Neural Network (CMNN) model to predict the shear strength of FRP-strengthened RC beams. • CMNN directly enforces monotonic relationships between variables through indicator vectors during training. • Compilation of 273 beams tests covering diverse strengthening configurations and parameters. • Automatically tunes CMNN hyperparameters using Bayesian optimization and 4-fold cross-validation. • Results show that CMNN provides more accurate predictions than design codes for different configurations. [ABSTRACT FROM AUTHOR]