Why I build AI agents like regulated programs

I spent fifteen years delivering enterprise programs in places where a missing control wasn't a bug — it was a finding, a fine, a name in front of a regulator. Wells Fargo, Bank of America, EY, Accenture. Risk, compliance, operational resilience, regulatory-driven change. Before that, two deployments to Iraq with the U.S. Army. The thread through all of it: when the stakes are real, you don't ship things that improvise.

Now I build AI agents. And the first instinct most people have with agentic AI is the one I've spent my career training out of teams: ship fast, let it figure things out, fix it in production. That works until it touches something that matters — a customer, a payment, a compliance boundary. Then "it figured it out" becomes "no one can explain what it did."

So I build agents the way I delivered regulated programs.

MAVIS is my production operations agent — live on Telegram and Slack, with persistent memory across conversations, plus email, calendar, file, and vision capabilities. She's not a chatbot bolted onto a workflow; she's an operator I can hand work to. But the part I care about most isn't what she can do. It's what she can't do without a gate.

Victoria is the team-facing side of the same philosophy. She runs inside a company's Slack — joining conversations, posting and updating team work status, asking team members for what's needed — and every channel is permission-scoped. Each team sees only what it should. That isn't a feature I added at the end; it's the foundation. Governance baked in, not bolted on.

Three things I refuse to skip.

Governance gates. An agent moving from sandbox to production crosses a controlled boundary, the same way a change does in a regulated bank. Nothing reaches production because it happened to work once.

Audit trails. Every meaningful action an agent takes should be reconstructable after the fact. If I can't answer "what did it do, and why," it isn't ready. That single question kills more bad ideas than any amount of testing.

Least privilege. Agents get the narrowest access that lets them do their job — scoped per channel, per team, per task. The blast radius of a mistake should be small by design.

Why this matters now.

Most organizations don't have an AI capability problem. They have an AI trust problem — leaders who can see the upside but can't get comfortable handing real work to a system they can't govern. That gap is exactly where my two careers meet. I'm not choosing between enterprise discipline and hands-on building. I'm insisting on both.

This site is part of the proof. It's multilingual, it's optimized for how AI search actually reads the web, and it has MAVIS embedded as a live voice and text agent — designed, built, and deployed by me. If you'd rather see the work than read about it, talk to MAVIS, or book a call.