Independent AI consulting · San Diego
I test AI claims before believing them. In December 2025 I built a benchmark that put seven language models through JSON extraction, mathematical reasoning, and executive summarization against fixed thresholds. One passed, and Towards AI published the framework and results. That is the habit I bring to every engagement: measure first, then advise.
Behind that sits a decade of teaching people to use complicated tools. Seven years as a graduate instructor at UC Santa Cruz (twenty-plus courses, more than a thousand students across every level of technical comfort). More recently, I onboarded and trained over two hundred distributed professionals for an AI evaluation firm, cutting performance variance across the group by 40 percent.
Earlier, I managed twenty-plus institutional accounts for a B2B education platform, including Harvard and MIT, with a 95 percent renewal rate. I hold a Ph.D. from UC Santa Cruz, which in practice means I read what a vendor is actually selling rather than what the deck appears to say.
The practice is named in Latin, which is the doctorate showing. Arte mentis: by the craft of the mind. The craft on offer is mostly judgment, which means knowing what to measure, what a claim actually says, and what to skip.
Seven local models tested against fixed thresholds; one passed. Methodology and results published in Towards AI.
1,000+ students across 20+ courses at UC Santa Cruz, plus 200+ professionals trained for an AI evaluation firm.
20+ institutional accounts managed for a B2B education platform, including Harvard and MIT.
Trained to read primary sources instead of summaries. It turns out to be the right training for reading vendor decks.
Test first, then believe. I put AI claims through fixed thresholds before acting on them, and most don’t survive. When I benchmarked seven models, one passed. That isn’t cynicism about the technology. It’s the reason you can trust a yes when I give one.
AI shouldn’t do everything. It should do the tedious things. The work that eats your week without needing your judgment is what these tools are for. The point of handing it off is more time for the work, and the life, only you can do.
The fear of being left behind is rational. Panic buying isn’t. Real things are possible now that weren’t ten years ago. The advantage goes to the businesses that implement deliberately, with security thought through before the contract is signed, not after.
I’ve worked both ends. I run local models on my own hardware and build with them; I’ve also deployed Microsoft Copilot inside an enterprise apparel company. I know what the corporate stack actually delivers, and what a small, private, right-sized setup can do instead.
The smallest model that clears the bar. Compute is electricity and water, spent somewhere real. Right-sizing the tool is good engineering for your security and your budget, and it’s the version of this industry a California business should want to be part of. Often it runs on hardware you already own.
Diagnosis is free; treatment is billed. The first 30-minute fit call costs nothing. Every engagement after it is scoped in writing, with a fixed price, before any billable work begins.
Independent means independent. I take no vendor commissions, referral fees, or reseller margins. Advice is priced so that the only person paying me is you.
Few clients, on purpose. I work with a small number of owner-led businesses at a time, so every project gets my direct involvement rather than a process.