A neural network on python built from scratch for training 28*28 images of hand written digits.The model uses gradient descent method for cost function optimisation.It consists of 3-layer network with 25 units in the hidden layer.The network is not yet optimised for training huge data sets.However it works fine on small training sets.