Most people learn ML frameworks by importing them.
We built ours by hand, then made it hit harder.
On Windows? See the PowerShell steps.
Tensor to Transformer, one op at a time: every module builds on the last, nothing skipped.
No black-box dependency. Every operation traces back to raw NumPy you can read in one sitting.
Memory, performance, and scaling analysis built into the curriculum, not bolted on after.
Rebuild the perceptron, the CNN, the transformer: the actual turning points, not toy examples.