I'm learning Machine learning by rebuilding the ideas behind the tools I use.

The stack I’m learning and building with as I turn ML fundamentals into small, understandable projects.

• Rebuilding a scalar automatic-differentiation engine in Python
• Understanding backpropagation instead of treating loss.backward() as magic
• Learning PyTorch through small experiments and classification projects
• Documenting the concepts, mistakes, and code as I go
• Working toward building a GPT-2-style model from scratch.

A small scalar autograd engine inspired by Karpathy's micrograd.

It currently includes:

• computation-graph tracking;
• reverse-mode automatic differentiation;
• arithmetic operations and tanh;
• hand-written gradient checks;
• a readable implementation designed for learning.

I am currently working on:

• comparing my gradients with PyTorch;
• adding neural-network building blocks;
• training a small MLP on XOR;
• becoming more comfortable writing Python without copying solutions.

I care more about understanding the mechanism than collecting libraries. My repositories are experiments and checkpoints from that process.