Research

Samir Awuapara

Solo researcher. Building neural network architectures from scratch in C++, Metal, and CUDA. No frameworks, no PyTorch.

Projects

Spark

A novel neural network architecture built from scratch in C++ and Metal. No frameworks, no pre-existing building blocks. Trained on a single M1 Pro laptop in under 20 minutes.

2.92

Best Val BPC

392

Training Runs

207

Merged PRs

20 min

Best Run Wall Time

C++ / Metal / CUDA Character-Level LM WikiText-103

Organic Cache Locality in Sparse Neural Networks

Dynamically grown DAGs self-optimize memory layout. 77-89% utilization, zero preprocessing.

Apr 2026

Dynamic Topology: Growing a Sparse Network from Scratch

39K → 3M parameters via dynamic growth. A self-assembling sparse DAG trained on a laptop in 20 minutes.

Apr 10, 2026

Atomic-Free Backward Passes on Metal

Eliminating atomics sounds like a win. On Metal, it's a 2.8x slowdown.

Apr 10, 2026

Metal Compute Shaders for Sparse ML

Zero literature exists on Metal ML kernels. Here's what we learned.

Interested in our research?

We are open to collaborations with researchers and institutions working on related problems.

Get in Touch