Not the latest AI guru. 30 years of combined engineering discipline. Now applied to agentic AI — so production-ready code stays a craft, not a shortcut.
[ what's agentic AI? ]
Not a chatbot that answers one question and stops. AI that takes multi-step action — looking things up, using tools, making decisions — to actually finish a task, with a human checking the work instead of doing it.
- Your team picked up ChatGPT, Copilot, Claude Code, Codex. Nobody picked up a system.
- Every person prompts differently. Every output is a coin flip.
- The pilot looked great in week one. By week three, nobody trusted it.
95%of generative AI pilots fail to deliver measurable business impact. Only 5% reach production.
MIT NANDA, "The GenAI Divide," 2025
Vibe-coding isn't software engineering. That's the gap — and it's why most pilots stay pilots.
Three decades engineering systems that had to work — not demo well once and fall apart in production. That same discipline now applies to agentic AI: building the memory, the verification, the structure that turns "we have AI" into "our team trusts what it produces."
No chasing this year's framework. Just principles that don't expire.
Harness engineering
The AI model is the engine. The harness is everything around it — memory, guardrails, feedback loops — that decides whether an agent survives contact with your actual business. Most pilots skip the harness. That's the part I build.
Software engineering isn't just code — it's a lifecycle: requirements, review, testing, deployment, feedback. Agentic AI is brilliant at the doing part of that cycle. It is not brilliant at knowing when it's wrong. Treat it like a new employee: capable on day one, but it still needs onboarding, guardrails, and someone checking its work until it's earned trust. Most teams skip that part entirely.
- "Is It Actually Working?" Call A working session tracing exactly where your AI's doing isn't backed by a checking — and whether the fix is a quick prompt change or a missing piece of the lifecycle.
- Build If the call finds real gaps, this is where they get engineered out — building the onboarding, guardrails, and feedback loops that turn a fast but unsupervised AI into a trusted part of the team.
Stop hoping the next prompt works. Find out what's actually broken.
Book the "Is It Actually Working?" Call