Systems study

RawTorch

RawTorch has two tracks. A NumPy-only library (TenTorch) implements tensors, automatic differentiation, and basic neural-network layers from scratch. Ongoing work in the CUDA path explores custom kernels, memory behaviour, and Python/C++ integration. It is a personal study project, not a production ML framework.

C++CUDAPython

View on GitHub

Highlights

Custom Tensor objects with an autograd engine and nn layers (Linear, ReLU, losses).
XOR training demos that force correct architecture and activation choices.
Experimental CUDA kernels aimed at understanding GPU execution, not shipping an ML product.
Bridge between high-level Python APIs and lower-level C++/CUDA behaviour.

Outcomes

Working NumPy-backed autodiff path suitable for small educational models.

Documented learning path into GPU programming and framework internals.

How to read this project

This demonstrates studying difficult low-level technical subjects. It is not a claim that the next role must be “ML systems specialist.” CUDA and framework internals sit beside software and infrastructure work as evidence of technical depth.