Higher-level but language-neutral abstraction layer for operators


I am trying my first bite of the operator development for MXNet, and I have learned that the backend is a place where new operators could be implemented, and have already successfully done something there. However, I wonder whether there is a possibility to place new operators in a higher abstraction layer where the interfaces of other defined operators are directly accessible, because it would be convenient sometimes to invoke other operators inside a new one. I know that ndarray.py might be a place for such operators but they would be bound directly to the Python language and therefore only accessible through the Python API. Is there such a layer that is higher-level than the backend layer but is still language-neutral?
What should I do, for example, when I want to implement an operator which calls reshape, broadcast_mul, etc?