<p>Bengali handwritten character recognition is a complex task due to the script’s wide range of characters, especially the compound forms that are frequently underrepresented or absent in standard datasets. To address this issue, a syn- thetic dataset was created comprising non-existing and rare compound Bengali characters, significantly enhancing the diversity and completeness of training data. This paper proposed BengNet, a novel hybrid convolutional neural net- work that integrates the strengths of both Inception and ResNet architectures. The Inception modules enable effective multi-scale feature extraction, while the residual connections from ResNet facilitate deeper learning and improved con- vergence. To evaluate its effectiveness, BengNet was compared against established pretrained models such as AlexNet and VGG16. On the test set, BengNet achieved a test loss of 0.1502, accuracy of 96.66%, precision of 97.01%, recall of 96.46% and kappa of 96.65%, outperforming AlexNet (86.55% accuracy), VGG16 (78.13% accuracy), ResNet50 (61.73% accuracy) and EfficientNetB0 by a significant margin. These results demonstrate the robustness and efficiency of BengNet, especially in recognizing complex compound characters. This work con- tributes to the advancement of script recognition for under-resourced languages by leveraging synthetic data generation and hybrid deep learning models.</p>