The notes behind the site.

Where showing a real artifact would expose confidential data, I rebuilt it at high fidelity instead. The work and the numbers are real; the customer specifics are invented.

Case pages mark these as they appear: a small line under the artifact saying it was recreated or redacted to protect confidential systems and customer data.

This site was built the way I build AI systems at work. Crews of AI agents drafted the copy and wrote the code; I set the direction, reviewed every page, and made the calls. Every number on the site traces to a receipt I keep: a claims ledger, the original artifact, or a documented measurement. One rule: no receipt, no claim. Before any page shipped, it passed three gates: redaction, honesty, and voice.

Delegating the typing doesn’t delegate the bar.

The "Ask my agent" chat on every page is a working agent I built: a small service that retrieves from these pages and answers in my voice, claiming only what the site claims. Every reply cites the pages it drew from. Some questions never reach the model at all: deterministic guardrails catch prompt injection, private matters, and out-of-scope asks, and answer them with rules instead. Before any change deploys, it has to pass a test suite, a golden set of graded questions, and adversarial cases that must fail closed.

It can be wrong, and it says so. When it is not sure, it hands off to the real me.

Instrument Serif for display, Instrument Sans for text, Fragment Mono for the working-file chrome, all self-hosted. The site is static HTML, CSS, and a little JavaScript: no framework, no build step. Hosted on GitHub Pages.