How to properly apply stochastic gradient descent to simply neural network?

Question

Im trying to write NN from scratch to better understand what is going on underneath Keras API. And now I have got problem with applying gradinets to my loss. I believe there