Profile / Denver, Colorado

About

Twenty years of engineering. Five cities. Four kids. A habit of ending up where being wrong is expensive.

I build AI systems for companies that are not allowed to guess.

At work today

The day job is Antech, the veterinary diagnostics business inside Mars, where I lead the move to agentic AI engineering. That phrase means something specific there. Agents read production, review pull requests, diagnose live incidents against the database and the monitoring, and draft the change-control paperwork that gates a deploy. Then a named person approves the write.

They have already worked a real outage end to end: read-only diagnosis, a drafted root cause, a drafted emergency change record, and a human who owned the fix and shipped it. The part I care about most is the plumbing underneath, a reviewed catalog of skills that turns one senior engineer’s judgment into something the whole team can load, with a merge gate no model can bypass. The case study has the numbers and the methodology behind each of them.

Earlier this year I ran a two-month contract alongside that job, founding the AI function at a regulated commercial real estate bridge-lending fund. I built the skill marketplace, the review gate, the monitoring behind both, and a daily investor-research service that drafted outreach for a named human to send. Two services were live in production at handoff. That write-up includes the parts that did not work, because those are the parts a hiring manager should read first.

In public

The same method runs my own projects, which is where I test it with nobody’s compliance officer watching. Something like fifty repositories and a dozen public domains are kept current by agents that run overnight. They research, draft, and open pull requests; I hold the merge.

• freecapitalists.org rebuilds the Mises Institute’s Creative Commons library into something a person can read: 949 books, 1,520 hours of lectures, and twelve million words of lecture transcript, most of it newly produced.
• cheatsheets.davidveksler.com has 170-plus reference pages built from one written spec, with a commit history anyone can inspect.
• coloradofirearmswatch.org went from a paragraph of dictated intent to production in about a day and has run itself since.
• whopaysforai.org grades who bears the cost of the AI buildout.

I also keep two older communities online, the Mises community forum and wiki and the Objectivism Online forum and wiki. I have run a crypto wallet recovery service with my name on it since 2017. Traffic for all of it lands on one public dashboard, which became its own small project. The full list is at /projects/.

The path here

I got here sideways. I studied economics and political science at Texas A&M and stayed for an MSIS, which is odd training for distributed systems until you notice that both subjects are about what people do when incentives meet constraints. Then I spent my twenties moving: Dallas, New York, Shanghai, Atlanta, Denver.

Four of those years were in Shanghai as an architect at EF Education First, where I learned that a system’s real specification is whatever the team actually does at nine on Monday morning, and the document describing it is a work of optimism. I came home, ran technology at the Foundation for Economic Education for five years, and built Liberty.me.

Then I went into fintech and crypto and stayed most of a decade. Institutional custody at Vellum Capital, which I co-founded and still help run. A wrapped-token launch and, later, the recovery work at Celsius that pulled $4M+ back out of stranded creditor vaults. A world-first Ethereum royalty platform at Royalty Exchange, which Billboard covered.

Celsius went bankrupt, and I was a senior engineering manager there when it did.

Here is the whole path in one table, since the prose skips around.

The overlaps are intentional. Vellum is a founder role I have run alongside operating jobs since 2018, and the lending-fund contract ran alongside Antech the same way. Education: MSIS and a BS in Economics and Political Science, both Texas A&M, plus General Assembly’s Data Science Immersive.

One opinion

Most enterprise AI programs fail because someone decided to fine-tune a model on company data. Fine-tuning bakes a snapshot of the business into weights that nobody can audit, diff, or roll back, and six weeks later the snapshot is stale. Retrieval against the system of record is cheaper, updates the moment the record updates, and leaves a citation a compliance officer can open. I have watched a team burn a quarter on a fine-tune that a decent index would have beaten in a week. I will argue about this at length if you let me.

Outside work

I train Brazilian jiu-jitsu, badly enough that it stays interesting. I have four kids. I live in Denver and expect to keep living in Denver.

What I want next is a job where building the AI capability is the job, at a company where being wrong costs money or hurts someone. If you are not sure whether that describes your company, the fit assessment will tell you in a couple of minutes, including when the answer is no. If it does describe you, email me at heroic@gmail.com.