Brain tumors pose a critical global health challenge, particularly in resource-limited regions like Bangladesh, where diagnostic tools must be precise and accessible. This work shows a new type of hybrid deep learning framework for classifying brain tumors (glioma, meningioma, pituitary) from T1-weighted contrast-enhanced MRI images, achieving an accuracy of 97.81%. Our novel approach integrates EfficientNetV2L, Vision Transformers, and a custom Window Attention mechanism, synergistically combining local and global feature extraction to overcome limitations of prior models, such as noise sensitivity and lack of interpretability. A tailored preprocessing pipeline employs Discrete Wavelet Transform and sharpening filters to enhance tumor visibility and robustness against MRI variability, crucial for diverse clinical settings. Mixed precision, gradual unfreezing, and a composite loss function that combines categorical cross-entropy and focal loss are all examples of advanced training procedures that deal with data imbalance and make sure that the model can generalize well. Grad-CAM visuals show areas of the tumor that are important, which builds faith in the model by giving clear explanations of its conclusions. Uniquely, the model is deployed via a Streamlit-based real-time decision support system, enabling radiologists to access predictions and visualizations seamlessly, particularly in low-resource environments. The framework achieves balanced precision, recall, and F1-scores above 97% across tumor classes, with minimal misclassifications. This work delivers a scalable, interpretable, and clinically relevant solution, advancing early brain tumor diagnosis and supporting improved patient outcomes in diverse healthcare settings.