Brett Wheeler

Independent · AI verification · fractional technical lead

You shipped it faster than you can check it.

I verify AI-built and AI-driven systems for founders and engineering teams. You get a written account of what holds up, what depends on a model behaving, and what nobody can currently account for — and, if you want it, the person who fixes that.

Work

Three ways to hire me.

Not long ago I spent a couple weeks evaluating three AI trading agents to find out whether a test I'd designed could even be run against them. It couldn't, cleanly, against any of them, and one of them was quietly reading two weeks of future prices. Nothing in their benchmarks showed it. That is the job: find out whether the thing in front of you can be trusted, and write down exactly why or why not.

01

Verification audit

For a founder with an AI-built product, or a team that added an AI feature, who needs to know what they've actually got before customers, investors or a regulator ask.

Two weeks with your repository and your running system. I read the code, run it, break it where it bends, and trace what depends on a model behaving the way it did last Tuesday. You get the readout in three parts: what's pinned to something checkable, what a human on your team still holds in their head, and what neither covers. Then a fix list in the order I'd do it.

Deliverable
Written report, 10–20 pages · fix list, prioritised · a 90-minute walkthrough · a go / no-go on whatever decision prompted the call
Terms
Fixed scope, fixed price, two weeks from repo access · NDA on request · price set before I read a line

$5,000

02

Evaluation harness

For a team shipping AI features without a way to tell a good release from a lucky one.

I build the thing that gates your releases. State isolated per call so one run can't contaminate the next. Scoring rules frozen under a tag before any real data, pinned by tests, so nobody can tune the threshold after the fact. Every report prints the commit it was scored under. I've built this for a published trading agent; the method moves to any system where a model makes a decision.

Deliverable
Containerised harness in your repo · frozen scoring rules with test suite · CI gate · one-page runbook your team can operate without me
Terms
Three weeks · your stack where it's sensible, Python and Docker where it isn't · I don't take the keys to production

$10,000

03

Fractional technical lead

For a company that needs the technical decisions made and owned, ten to fifteen hours a week, by someone who has shipped inside banks, insurers, exchanges and a few startups, and who will build rather than only advise.

Architecture. Which model, which vendor, API or self-hosted, and what it will cost you a month from now. Security and the questions an enterprise buyer or an investor's diligence partner will ask. The first engineering hires and how to tell whether the agency you're using is any good. I take decision authority inside an agreed scope, deliver in writing, and don't carry a pager.

Deliverable
Weekly written decisions log · architecture and roadmap documents · hands-on build time inside the retainer · investor and diligence support
Terms
From 10 hours a week · three-month minimum · scope and authority set in the statement of work · no on-call

From $6,000 / month

Method

How I work, written down.

Most of my recent code is written by AI agents under my direction. I say so in every repository and every paper, because the point isn't who typed it. The point is whether anyone can check it.

  1. One change per prompt.

    An agent gets one bounded task. I read the result before it lands. Scope errors get caught in the diff, not in production.

  2. Nothing merges unratified.

    Every commit is reviewed and signed by me. Conventional commits, reference trailers, the agent credited as co-author.

  3. Rules freeze before data.

    Any scoring, threshold or acceptance rule is tagged in git before it sees a real result. Afterwards it can be changed, but not quietly.

  4. Cold review.

    A second agent session with no memory of the first checks the work. Same model is fine; shared context isn't.

  5. Reports print their own provenance.

    Every output names the commit it came from. If you can't reproduce it, I haven't finished.

I'm writing this method up as an open specification. It's called OpenKoine, and it's on the final approach - landing soon.

Track record

Twenty years, in specifics.

Regulated and enterprise integration

  • Futures-hedging portal with position and risk calculation; authenticated API and flat-file integrations to CME and trading partners (industrial, Iowa)
  • Core-system storage migration from SQL to a document store, feature-flagged with dual-store sync, on an 11-person agile team (insurance, Omaha)
  • Multiple codebases carried from .NET Core 3.1 to .NET 6 inside a SAFe release train (agricultural lending, Omaha)
  • Project lead at a compliance and ethics-reporting software company (Denver)
  • Order-entry module embedded in HIMSA Noah, the platform most of the world's audiologists use (hearing technology)

Products shipped end to end

  • Slot-machine inventory and KPI system, requirements to client adoption, then re-engineered into a multi-tenant SaaS with billing and tenant administration
  • Three back-end subsystems for a SaaS eCommerce platform: adaptive import/export, job processing, CDN image offloading
  • Two subscription web products for faith and eldercare audiences, built on Next.js with AI coding agents (2024–26)
  • tashika.co: Next.js, Tailwind, shadcn, TypeScript, OKLCH design system, repository-driven essay publishing

Research

  • FinPerturb: containerised evaluation harness, four pilot series, 600+ scored decisions; source audits of FinMem, TradingAgents and FinAgent; preregistered design (OSF), archived code (Zenodo)
  • The Substrate Collapse, arXiv:2606.20882 (cs.SE)
  • OpenKoine, an open specification for working with memoryless agents under provenance and ratification (in progress)

Stack

  • C# / .NET, Python, TypeScript · SQL Server, Cosmos DB, Postgres (Supabase, RLS) · Azure, AWS, Cloudflare Workers, Docker · Next.js, Vue/Nuxt, Angular, WPF

B.S., Computer Information Systems Management, with honors. Résumé on request.

Contact

One email is enough.

hello@brettwheeler.com

What to send

  • A link to the repository, or a sentence about why you can't share one yet
  • One paragraph on what you shipped and what's worrying you
  • Which of the three you think you need; I'll tell you if I disagree

I reply within two working days. If I'm not the right person I'll say so and, where I can, name who is.