Work

Production AI fails at the seams — between the model and the data, between the demo and the operating reality. My work lives at those seams.

What I build

Agent architectures that survive contact with production. Memory designed as a data modeling problem rather than a retrieval afterthought; typed records with provenance and forgetting policies; evaluation harnesses that tell you an agent is drifting before your users do; failure-mode analysis as a first-class design activity, not a postmortem one.

Data foundations agents can actually stand on. An agent is only as trustworthy as the data it acts on. I design the lakehouse patterns, data models, and governance structures that give AI systems something true to work with — which is where most "AI projects" quietly reveal themselves to be data platform projects.

Teams that can run the system after I leave. Every engagement transfers capability: architecture decision records the team can extend, operating runbooks, and structured training built on progressive complexity — so the system's survival doesn't depend on my continued presence.

How I work

Top-down: the shape of the whole is decided before any part is developed in depth. Verification-oriented: generated content, vendor claims, and first drafts all get checked. Consistency as a feature: the same concepts get the same names and structures everywhere. The longer version is on the About page — and this site itself — built spec-first as a dependency-free system (see specs as forcing functions) — is a small working example.

Case studies

In the spirit of the garden, this section is honest about its maturity: the write-ups below are being tended now and will be published as they complete.

• Reference build: agent memory over a lakehouse

  An enterprise-grade agent with a properly typed memory layer — fact, preference, episode, reference — over lakehouse data foundations, with provenance and forgetting policies designed in from the start. Full architecture case study in progress.

  Agentic AI
• Reference build: evaluation & observability harness for agent systems

  The instrumentation an agent system needs before anyone should trust it: behavioral evals, drift detection, and the operational runbook that goes with them. Write-up to follow the build.

  Agentic AI

Wrestling with the gap between an AI demo and a production system? Email me and let's talk.