See how AI models actually work — shape by shape.
Hand-drawn, interactive walkthroughs of real architectures. Zoom into the diagram, follow every matrix multiplication, then run the code in Colab.
Lessons
LLVM & Compiler Architecture
Why compilers split into front end, IR, and back end — and how that one decision lets n languages target m machines without an n×m explosion of compilers.
compilersllvmirsystems
Llama 3 Mini, from scratch
A tiny Llama-3-style transformer (~40M params) traced shape-by-shape, then trained on TinyShakespeare. Every matrix, every dimension.
transformersllamafrom-scratchattention