This paper deals with the bow optimization of a RoRo cargo ship to reduce full-scale resistance in calm water. It employs a multi-fidelity surrogate model integrating high- and low-fidelity potential flow solver solutions. Adaptive sampling, driven by a dynamic lower confidence bound strategy, determines both new sample locations and fidelity levels. Design-space dimensionality reduction via parametric model embedding enhances optimization efficiency. The optimized design is validated through full-scale RANS simulations in OpenFOAM, ensuring accuracy and reliability. This study highlights the advantages of combining multi-fidelity models with active learning strategies for hydrodynamic design optimization. The ability to dynamically select both the sampling location and fidelity level enables significant computational savings without compromising accuracy. Furthermore, validation with RANS simulations demonstrates the reliability of this approach in predicting performance improvements, paving the way for broader applications in naval engineering.