Creating a library for symbolic algebra in Nim is pretty cool, but you don't need to follow the design of Python where the expression tree is constructed at run-time through operator overloading. In Nim, you can directly work with the AST-form of arbitrary formulas at compile-time by creating a macro that will take a block of code. You can also use the macros.getImpl family of magics to obtain the code of existing functions and transform it in arbitrary ways.
This approach will give you more power - not only you will be able to differentiate functions, but you'll be also able to turn the computed derivative into generated code that will be compiled with full optimisations by the C/C++ compiler. Gradient descent with back-propagation can be applied to many fields and it can greatly benefit from a capability like this. ArrayMancer is a famous library that already provides something similar: [https://github.com/mratsim/Arraymancer/tree/master/src/autograd](https://github.com/mratsim/Arraymancer/tree/master/src/autograd)
