Writing

MCP shipped its first deliberately breaking spec revision. My agent factory runs on an in-house MCP SDK — so I wrote the gap analysis, filed the roadmap as factory backlog, and Mills started rebuilding the ground it stands on. Receipts included.

How I run LLM agents against regulated healthcare payment data without losing sleep: a three-tier trust ladder, kill-tests before promotion, and guardrails that live in infrastructure instead of prompts.

Four defect classes that healthcare interchange parsers accept silently, with synthetic samples, real linter output, and the pre-send gate that catches them.

The sequel to Five Voices: in one day, a config audit became three production promotions (prefix caching, fp8 KV, a doubled 703K-token pool), one spectacular-but-vetoed 4.35× decode win, two remembering psyches running side by side on a single consumer GPU — and five real bugs caught by a twin-lane canary loop before production felt any of them. Here is the method, the numbers, and everything that broke.

We rebuilt our Jungian multi-agent lab so each archetype carries its entire life — every situation, everything heard, everything said — as an append-only journal that fills a 228K-token context window on a 24 GB Radeon. The prompt is shaped so the whole life is one reusable prefix; when the window fills, the agents dream their oldest memories into a durable core. Here is the architecture, the numbers, and what broke along the way.

A served model can load at 96K context and still be useless past 64K. Loading is not the same as staying coherent. Here is how we mapped the exact cliff with a progressive needle-in-haystack bench — and why the limit was the model, not the GPU.

How I would roll out AI-assisted development on a team that has not standardized: what to do in week one, what to earn the right to argue about later, and what almost always goes wrong.

A short, adoptable AI usage policy for engineering teams: what to put on the page, what to leave off, and why the policy matters less than the habits it makes explicit.

What it took to get Gemma 4 E4B serving cleanly on Radeon through FlexInfer: a stable TRITON lane on a 7900 XTX, an experimental TurboQuant long-context lane on a second node, and the GPTQ pipeline work still underway.

AI-assisted development is useful leverage, but only if you convert borrowed competence into real judgment before the support becomes a dependency.

Field notes from building and operating a small private GPU platform with Harvester, K3s, and a GitLab -> Flux delivery loop.

A deep dive into optimizing Canvas 2D and Three.js visualizations for Kubernetes dashboards, covering algorithmic complexity, memory management, and GPU-efficient rendering patterns.