Start typing to search notes, projects, and books.
I work on machine learning systems, from implementing neural networks and automatic differentiation to reproducing sequence models and studying how models train, reason, and generalize.
Exploring: Deep Learning · LLMs · Training Dynamics · RLHF
A practical-scale reproduction of the 2014 sequence-to-sequence architecture, with explicit LSTM gates, controlled source-reversal experiments, and deliberately reported failure modes.
Four small-scale CPU runs reported test BLEU of 1.248–2.715. Source reversal did not consistently help.
Learning extensions to Andrej Karpathy's micrograd, with input-shape fixes, gradient regression tests, a backpropagation walkthrough, and reproducible command-line training.
Tests cover shared operands and gradient accumulation. The scalar engine has no tensors or GPU acceleration.
A NumPy-only character language model that implements the recurrent forward pass, BPTT, gradient checking, Adam, checkpointing, and controlled sampling by hand.
The continuation run reported 2.1348 validation loss per character; generated text still lacks consistent coherence.
Learning-rate decay reduces the effect of noisy late-stage updates while preserving larger steps early in training.