Malaria remains a significant global health concern, particularly in regions where diagnostic infrastructure is limited. Recent advancements in artificial intelligence (AI), especially Convolutional Neural Networks (CNNs), have enabled the development of automated diagnostic systems capable of detecting malaria with high accuracy. In this study, we evaluate two custom-designed CNN architectures for the binary classification of parasitized versus uninfected red blood cells using the NIH (National Institutes of Health) Malaria dataset. The architectures, deliberately minimal in complexity, are tailored to resource-constrained environments, featuring only three convolutional layers and varying configurations in convolutional filter width and dense layer size. Though the architectures are simple by design, it is believed to be novel, especially in regards to its application to detect malaria affected human red blood cells in POCT (Point-of-Care Testing). We further investigate the impact of dynamic, on-the-fly data augmentation on model performance and generalizability. All models achieved comparable accuracy (~94%) with a very limited number of training epochs; the incorporation of augmentation proved beneficial in mitigating overfitting and stabilizing validation metrics. Notably, the leaner architecture, despite having significantly fewer parameters, matched the performance of its more complex counterpart and demonstrated improved recall for infected samples when trained with augmentation - an essential trait in clinical screening to minimize false negatives. Our findings underscore the practicality of deploying lightweight CNN-based systems for POCT malaria diagnosis. We also discuss the broader challenges in this domain, including dataset generalizability, the need for multi-class classification and the integration of interpretability tools.