$ tail -f ./notes

Notes from production

Long-form breakdowns of systems I've actually shipped. No listicles, no hot takes.

7 posts

PERF2026-09-14 · 9 min read

255ms to 68ms: a --version fast-path for KiroCrew

Porting Claude Code's 12ms --version trick to an open-source Python CLI: two argv guards, six review rounds, one accidental PR closure, and a maintainer bot that finished the job.

read_post →

STREAMING2026-09-10 · 19 min read

An S3 OutputStream needs a success signal

A failed export can leave a successful S3 object. Explicit commit/abort semantics, bounded buffering, and failure behavior, verified against real S3.

read_post →

PLATFORM2026-08-25 · 15 min read

Unused permissions, priced in dollars

A license-saving feature end to end: live-computed findings, one action that becomes removal tasks, and action records engineered to survive the nightly data refresh.

read_post →

INFERENCE2026-02-10 · 14 min read

Route to the machine that already knows the conversation

Most AI serving cost is re-reading conversations a computer has already seen. Nostos sends each request where its past already lives. Plus a public arena for racing routing strategies.

read_post →

ML2026-01-18 · 13 min read

A language model from scratch, three times

The same transformer at three levels: PyTorch with no shortcuts, dependency-free C++ with tests green, raw GPU code in progress. What changes each time you rebuild it.

read_post →

AI SAFETY2025-10-15 · 13 min read

SoD for AI agents: toxic combinations in autonomous systems

Five risk types across agents, owners, chains and credential groups. AGENT_ONLY to INVOKER_COMPOSITE, in a deterministic three-pass engine.

read_post →

PERF2025-09-01 · 12 min read

From ~20 hours to minutes: a compliance check, re-examined once

A compliance check re-examined the company hierarchy once per rule. Resolving it a single time collapsed a worst-case production run into minutes.

read_post →