← All work

ML · From scratch · 2025

GraphFlow

A small graph-learning library built in plain NumPy, where the parts that usually stay hidden (autodiff, message passing, graph generation, the GNN layers) are all right there to read, and still fast enough to benchmark.

Diagram of GraphFlow autodiff and graph neural network components.
The NumPy autodiff and GNN components.

Problem

The big graph libraries are excellent, but they hide a lot of what is actually happening underneath.

Approach

I wrote the tensor operations, optimizers, graph kernels, and a few GNN architectures myself in NumPy, with reproducible validation workflows on top.

Outcome

Something simple enough to learn from but complete enough to benchmark: 40% lower latency, 30% better memory efficiency, and over 90% unit-test coverage on the differentiable message passing.