During the development of an AI model for train rail crack detection, we faced several challenges. First, we had a small dataset for model training, which could lead to overfitting. To address this, we employed data augmentation techniques, generating additional samples through random transformations to increase dataset diversity and improve generalization. Initially, the model's accuracy was only 50%. We improved this to 81% through various steps, including data augmentation, comparing different base models such as CNNs, ResNet, and InceptionNet, and implementing regularization techniques like dropout and L2 regularization. Additionally, we faced resource constraints for exhaustive hyperparameter fine-tuning. To overcome this, we used efficient hyperparameter tuning methods such as random search and Bayesian optimization, along with cloud-based solutions for parallel experiments. These strategies allowed us to optimize the model's performance within our resource limits. Despite these challenges, our efforts resulted in a robust AI model capable of accurately detecting rail cracks and a user-friendly web app for predictions. The project required continuous experimentation and innovative techniques to achieve the desired outcomes.