← Index

Writing

Notes from nights & weekends — mostly inference optimization, post-training, and things I got wrong before I got them right.

2026

• The KV cache is not your bottleneck Profiling a 70B serving stack and finding the time somewhere else entirely. 07/18
• Rejection sampling beat my reward model Six weeks on a per-span RL controller, and the two-stage baseline that replaced it. 05/04
• Shipping AI on a railway What "mission-critical" does to your eval harness, your rollout plan, and your sleep. 03/22

2025

• Speculative decoding, honestly measured Acceptance rates look great until you plot them against batch size. 11/09
• A byte-movement model for small accelerators Why FLOPs stopped predicting latency on the tiny TPU, and what did. 08/13

---