In this work, the automatic recognition of handwritten digits was treated using a classical and modern machine learning approach; both based on neural networks. First, we present a simple softmax regression model and achieved a recognition accuracy of 91.91%. Then, we implement a convolutional neural network model with a custom KNet architecture and achieved an accuracy of 99.13%. In this work, we make use of make use of the MNIST Dataset which comprises of cleaned train, test, and validation categories of hand-written digits as grayscale images.