Writing
Reports and notes from experiments, with the data behind them.
- LoopGYM: RL on Bugs My Agents ShippedOctober 16, 2026
GRPO post-training of a small coding model on Exercism and on bugs my own coding agents shipped, plus the training stack to run it on two 3090s. Most of each batch turned out to carry no learning signal.
- Memory vs. Weights for CTF AgentsOctober 15, 2026
Can an agent get better at capture-the-flag challenges just by writing down what worked? Six learning arms on one frozen agent, compared against QLoRA fine-tuning on the same verified experience.
- Muon: Accuracy or Shape?October 15, 2026
Built a certified family of Newton–Schulz polynomials and ran a budget-matched study on two 3090s to see whether Muon works because it orthogonalizes well, or because of the spectral shape it leaves behind.
- Speculative Decoding for Agent WorkloadsOctober 6, 2026
How much of a coding agent's output can a small model predict, and does that make generation faster? A LoRA drafter overlaps with the target on 71% of the probability mass and makes a 1.5B target 1.58× faster on CPU.

- Local Multi-Agent LoopSeptember 28, 2026
Ran a two-agent loop and tried to optimize context and task observability both internally between agents and externally for reviewers.

- Distributed Rate Limiter2024
A token-bucket rate limiter backed by Redis Lua scripts, designed to handle 50k req/s with sub-millisecond overhead and correct behavior under network partition.
- Search Ranking Pipeline2023
A learning-to-rank pipeline that replaced a hand-tuned BM25 baseline, shipping an offline evaluation framework that made iterating on ranking models safe and fast.