moe chabot

Construct a minimal deep learning framework from scratch, layering a lazy numpy-backed buffer, a reverse-mode automatic differentiation engine, a tensor API, and neural network primitives. By the end you train a small MLP on a toy digit dataset, mirroring how real autograd libraries like tinygrad work under the hood.

Reimplement the original encoder-decoder Transformer end to end in PyTorch, from token vocabularies and sinusoidal positional encodings through multi-head attention, label smoothing, Noam scheduling, and beam search. By the end you will have a working seq2seq Transformer training loop and inference pipeline assembled from first principles.

A suite of plugins for legal workflows