Writing

neoKernel: a research loop for faster inference

Self-improving harness for Qwen3 inference on H100. A model proposes a bounded kernel or engine change; the loop guards the source, runs correctness checks, benchmarks against native, and keeps or reverts the candidate. 27 trials, ~700-800 tok/s.

Training a 30M-param GPT from scratch

Multi-head self-attention, positional encoding, training on tinyStories with JAX / Optax / Orbax on a T4. What broke, what worked, what I'd do differently.