Projects

A single-turn word-ladder puzzle with a password constraint, built to be trivially verifiable for RL training of reasoning models. Includes a generator, graded difficulty tiers, frontier-model baselines, and a playable version in the browser.

A minimal GPT-2 implementation in NumPy with a KV cache, written to show exactly what gets cached during generation and why it makes decoding faster.

TensorFlow implementation and results of "An Introduction to Deep Learning for the Physical Layer": learning a transmitter and receiver end to end as an autoencoder over a noisy channel.

PyTorch implementation of "A Deep Multi-Level Network for Saliency Prediction".