Independent AI Assurance
When you put AI in front of customers, capital, or regulators — especially the autonomous, agentic systems now taking real actions on your behalf — a single undetected failure can cost more than your entire AI budget. I find it first — through evaluation, adversarial red-teaming, and governance hardened in regulated finance — and hand you the evidence to deploy with confidence.
Every team can ship an AI demo. Far fewer can tell you, with evidence, where it fails and how badly.
A model that performs beautifully in a sales demo can hallucinate, leak data, or quietly drift the moment real users — and real adversaries — arrive. The gap between "looks impressive" and "behaves reliably" is exactly where reputations and capital are lost.
I close that gap. I treat your AI the way a regulator, an attacker, and your most demanding customer would — then hand you the evidence to act on, in language your board and your engineers both understand.
What an engagement delivers
Ways to work together
Every engagement begins with a single assessment and a system stood up around your AI — then continues as ongoing assurance, at the altitude you need.
A fixed-scope diagnostic of one AI system — what it gets wrong, and how badly.
Set the standard, then build the assurance pipeline that enforces it — and hand your team the keys.
Continuous, automated assurance running against your AI — with my oversight on what actually matters.
Your fractional head of AI assurance — a standing seat at the table.
Most clients begin with an assessment, build their pipeline, and settle into Assured — a select few step up to Embedded. That mix is the practice.
Agentic AI assurance
Autonomous agents don't just generate text — they call tools, move money, execute code, and chain decisions across steps. A jailbreak used to be an embarrassment; in an agent, it's an unauthorized transaction. I assure the whole agentic stack against the failure modes built for systems that act, not just answer.
Hallucination, data leakage, prompt injection — the request-and-response layer.
Goal hijack, excessive agency, memory poisoning, rogue behavior — the autonomy layer.
Tool misuse, privilege abuse, supply-chain compromise — the connectivity layer.
Tested against the frameworks built for AI that acts — the agentic and tool-layer risk taxonomies, not just the model.
Selected work
Engagements are confidential; the patterns are not. Each of these is the kind of failure that stays invisible until it's expensive.
Engaged by — Head of Risk, lending fintech
A fintech was weeks from launching an LLM that pre-screened loan applications. The adversarial assessment surfaced a prompt-injection path that could coax the model into approving applications it should have flagged — invisible in every internal demo.
Engaged by — VP Engineering, health-tech platform
A patient-intake assistant looked safe — until specific phrasings caused it to surface protected information it should never have returned. The red-team found and reproduced the path, then built it into a test suite that now runs on every model update.
Engaged by — CTO, B2B software company
An AI support agent was confidently inventing refund terms and commitments. There was no measurable bar for "acceptable," so no one could say whether it was getting better or worse between releases.
The math
Premium clients don't ask "what does it cost." They ask "what does it cost me to be wrong." Here's the trade you're actually making.
The work is the cheapest insurance you'll buy this year —
and the only kind that makes you faster, not slower.
Why me
I spent my career hardening software in regulated banking — where a missed defect isn't a bug ticket, it's an audit finding, a fine, or a headline. That environment teaches a specific discipline: assume nothing, prove everything, and write it down so it holds up under scrutiny.
Away from the screen, I build furniture by hand. A mortise-and-tenon joint either holds or it doesn't — there is no convincing demo that hides a weak joint. The same instinct drives how I evaluate AI: the work that matters is the work nobody sees until it fails.
Most AI testing is run by people who learned to test on AI. I bring a career of testing systems where the stakes were already real, and apply it to the newest place those stakes now live.
How I work
Every engagement runs the same disciplined sequence. You always know where we are and what you're getting.
We agree on the behaviors that matter, the risks you can't accept, and the bar your AI has to clear — before a single test runs.
I evaluate and attack the system across the dimensions that matter — accuracy, safety, leakage, robustness — and capture exactly how it responds.
Results become scored, reproducible, versioned records — not anecdotes. You can re-run them on every release.
Prioritized findings, recommended controls, and the governance to keep the bar from slipping after I'm gone.
In their words
He found in three weeks what our team had walked past for six months — and made it impossible to ignore.
The first time anyone handed us a number about our AI that we could actually defend to the board.
We brought him in to check a box. He changed how we ship AI.
About
An independent AI assurance practice — elite evaluation, adversarial red-teaming, and governance, led by a quality engineering veteran out of the regulated-finance world.
Most clients run on a continuous assurance program backed by tooling I've built; a select few hold an embedded advisory seat. Either way, the judgment is mine — the system does the heavy lifting, not a junior bench. Based in the Dallas–Fort Worth area, working with clients anywhere.
If you're putting AI in front of customers, capital, or compliance, and you need to know — really know — that it holds, that's the conversation I want to have.
Start a conversation
Engagements begin with a direct conversation — no forms, no funnel. Tell me what you're deploying and what keeps you up about it.
Assurance subscriptions open · a few embedded seats each quarter
Email me directly