The article explores the use of autoencoders to tackle the challenge of credit card fraud detection, a significant issue for banks, with fraud accounting for about 6.8% of international transactions despite efforts to reduce it. Using a Kaggle dataset of European transactions from 2013, where fraudulent transactions represent a mere 0.172% of the total, the study outlines the challenges posed by class imbalances and the irrelevance of traditional evaluation metrics like accuracy. The data undergoes transformations and scaling, and an autoencoder model is trained on normal transactions to learn implicit representations of normality, which helps in distinguishing fraudulent transactions during inference. Precision-recall tradeoffs are analyzed, revealing an optimal threshold for fraud detection, and the autoencoder's effectiveness is highlighted through visualization techniques like t-SNE. The article emphasizes the importance of retraining on misclassified samples to improve model performance and demonstrates the utility of tools like Comet for experiment tracking and data analysis, ultimately achieving a precision, recall, and F1-score of 0.91, with further potential improvements suggested through detailed data examination.