// AI architecture & consulting

Your pilot works.
Production is the other 90%.

I take AI agents the rest of the way. Architecture, deployment, and an evaluation set that tells you it still works in March. Then I hand it over and go.

Start with a $2,500 one-week audit Read the Field Manual first

The demo was never the hard part.

Getting an agent to work once is a weekend. Getting it to work reliably, under audit, on data you do not control, after the model changes underneath you, is the job.

Most teams are not stuck because they picked the wrong framework. They are stuck because the thinking got skipped. The tool was bought before the problem was written down. Then nothing downstream of the demo was built: no evaluation set, no rollback, no way to tell whether it is working except by asking someone who is guessing.

  • 10%

    In any given business function, no more than a tenth of organisations report actually scaling AI agents. Only 39% report any EBIT impact from AI at all.

    McKinsey, State of AI, Nov 2025 · n=1,993 source

  • 40%+

    Share of agentic AI projects Gartner expects to be cancelled by the end of 2027, on escalating cost, unclear value, and inadequate risk controls.

    Gartner, Jun 2025 source

  • 17%

    Organisations that continuously monitor agent-to-agent interactions. Only 38% monitor AI traffic end to end at all. You cannot debug what you never instrumented.

    Cloud Security Alliance, Apr 2026 source

If your agent has not been built yet, start at the audit. If it has been built and nobody can tell you whether it is working, start at the acceptance criteria.


Start small. Decide from evidence.

One week. One fixed price. You own everything it produces, whether or not we work together again. If you continue, the fee credits forward, and the larger work is scoped from what the audit found.

02

Agent Pilot Sprint

  • 4 to 6 weeks
  • scoped from the audit

For teams with an approved use case who need it built properly the first time.

  • A working agent. Against real tools and real data, not a scripted demo.
  • State, memory, and recovery. Designed for the failure modes you will actually hit.
  • An evaluation set and a pass criterion. Agreed before the build starts, so "is it working?" has an answer that is not a hunch.
  • Tracing from day one. On OpenTelemetry GenAI conventions, so you are not locked to one observability vendor.
  • A deployment path. Repeatable releases, a documented rollback, and a runbook written for whoever operates it after me.

Outcome: Something in front of real users, instrumented well enough that the next decision has data behind it.

Built with LangGraph · LangChain · Microsoft Agent Framework · Google ADK · MCP · Claude Code · OpenCode

  • Multi-agent design
  • Tool integration
  • Evals & acceptance
  • Observability

03

Embedded AI Architect

  • monthly
  • scoped to your team
  • cancel monthly

For teams with capable engineers who want someone who has already made these mistakes sitting alongside them.

  • A fixed number of days each month. Named up front, on your architecture and your review queue.
  • Design review and code review. Including the unglamorous decisions: model routing, cost ceilings, context strategy, what to cut.
  • Your engineers learn the method. Not renting it from me. These engagements are meant to end.
  • No leverage held over you. No minimum term, no exit fee.

Outcome: Your team ships faster and needs me less each month.

  • Architecture review
  • Design partnership
  • Team capability
  • Cancel monthly

// Applied across all work

  • Governance & compliance
  • Documentation & SOPs
  • Knowledge management
  • Agent observability
  • Performance monitoring
  • Process design

You keep the keys.

Small business tech has a dependency problem. A vendor builds your site on their platform, registers your domain in their account, and holds every login. When the relationship sours or they get complacent, you are a tenant in your own digital property.

Here it works the other way. The domain lives in an account you control. The hosting is yours. The login to edit your own site belongs to you. Stop working with me tomorrow and you lose nothing you paid for.

You should also be able to get a change made without phoning me. Anyone you choose to pay can work on it.

A family flooring contractor in New York City went live on 26 October 2026 on hosting, a domain, and a repository the business owns. The previous site was archived and handed over. Read what that took.


Shipped, and checkable.

Every client is described by sector rather than named, under NDA or until that client agrees in writing to be named. What is described instead is the architecture, the constraint that shaped it, and the part I owned. Each one is labelled with what it is. One is a hackathon build, and it says so.

Small business · Web platform migration

Shipped, client not named

A family flooring contractor off WordPress and onto something it owns

WordPress to a statically built Astro site, 297 legacy URLs preserved, launched 26 October 2026.

297 legacy URLs preserved, 1,440 redirect assertions green, 32 of 32 CDN edge checks serving the new site

  • Astro
  • GitHub Actions
  • Migration
  • Technical SEO
  • GA4

Education technology · Collaboration

Shipped, client not named

An adaptive learning engine that has to be right about a child

A Next.js homeschool platform whose grading engine adjusts tomorrow's difficulty per concept, and reports to a parent in plain English.

  • Next.js
  • Adaptive learning
  • Content engine
  • Evaluation

Global logistics & enterprise IT · AIOps

Enterprise delivery, under NDA

Agentic operations platform for SAP and ServiceNow

Alert triage across SAP and ServiceNow, traceable end to end, with a human in the loop on anything consequential.

  • Microsoft Agent Framework
  • OpenTelemetry
  • RAG
  • SAP
  • ServiceNow


Ready to ship something that matters?

Enterprise AI, a small business moving off a platform it does not own, or learning the method yourself. Start with a conversation.