The Lights Are Back On

TL;DR — I stopped writing here in 2019. I didn’t stop building. I’m back, and the blog itself is the first artifact.

Where I went

The last post on this blog was an AWS Kinesis demo in November 2019. Then work ate the writing.

In fairness, it was the good kind of eating. Over those years I got to build the guts of a regulated investment platform: a secondary market for private securities, escrow and custody integrations, the cap-table machinery, and more recently the part I’d have called science fiction in 2019 — production AI systems whose output gets reviewed by compliance officers and regulators, not just users.

Along the way I co-authored ERC-1450, an Ethereum standard for transfer-agent-controlled security tokens — the reference implementation is public and security-audited. Over a billion dollars of securities has been issued on that pattern now, which is the kind of sentence you don’t get to write about most side quests.

The writing never actually stopped, it turns out. It just went internal — hundreds of design docs, runbooks, and explainers on a company wiki. At some point I noticed that I’d quietly become the person who documents everything, and that almost none of it was visible outside a login page. This blog restart is me fixing that.

What happened to the blog

This site has run on WordPress since 2003 — before that I was on JRoller, which tells you something about how long I’ve been doing this. WordPress served me fine for two decades of occasional posting. But when I came back to it this week and tried to make it look and behave the way I wanted — real typography, syntax highlighting for code, a design I actually own — I kept hitting platform walls.

So I did the engineer thing. The blog is now a static site: every post is a markdown file in a git repo, built with Astro, served from GitHub Pages. All 121 posts back to 2003 came along, at their original URLs. The 2009 Clojure posts have syntax highlighting now — they waited sixteen years for it.

I left the old posts up, unedited. Some of them are about RethinkDB and Maven and upgrading to Snow Leopard. That’s fine. They’re an honest record of what building software looked like at the time, and pretending otherwise would be revisionism.

What’s next

Two things, mostly.

mcpwright — a suite of open-source MCP servers I’ve been building that put public data inside AI agents: SEC EDGAR filings, U.S. Census data, more coming — all on github.com/mcpwright. Each one is small, typed, tested, and installable in Claude Desktop with one click. There’s a surprising amount of craft in making a good tool surface for a language model, and I’ll be writing about it.

The backlog — twenty years of internal writing has left me with a pile of things worth saying publicly: how we evaluated token standards, what production LLM systems in a regulated company actually look like, why a knapsack algorithm ended up allocating ad budgets, and the unreasonable effectiveness of writing things down.

If you’ve read this far: hello again. The pens are inked, the ukulele is roughly in tune, and the lights are back on.