emlgrad

emlgrad is a tiny scalar autograd toy with strong micrograd vibes.

The twist is that it only has one primitive math op:

eml(x, y) = exp(x) - ln(y)

This project is mostly inspired by:

• Andrej Karpathy's micrograd
• Andrzej Odrzywolek's paper, "All elementary functions from a single binary operator"

So the basic idea is:

What if a micrograd-style engine kept the same small DAG + backprop feel, but built its math by composing EML?

What's here

• emlgrad/engine.py: the scalar Value engine and backward pass
• emlgrad/nn.py: a very small micrograd-style MLP wrapper
• demo.ipynb: a notebook in the spirit of the original micrograd demo

Higher ops like log, +, -, *, /, **, and relu are all lowered into eml.

Notes

• complex-domain first
• principal-branch log
• branch cuts and log(0) are not hidden away
• some paths can still produce inf or nan

This repo is meant to stay small, readable, and a little weird.

Quickstart

This repo uses uv.

uv sync --group dev
uv run python -m unittest discover -s tests
uv run --group dev ruff check .

To open the notebook:

uv run --with jupyter jupyter notebook demo.ipynb