A decade of Ruby on Rails, and the last three years spent building agentic systems that actually ship. Currently Architect on the VA Appeals modernisation programme through Take2 Consulting in partnership with Booz Allen Hamilton, where I own technical direction across several platforms and led a team-wide AI tooling rollout.

Before that, seven years building a HIPAA-compliant behavioral health EHR. Two regulated domains, learned to the level where the details matter.

Forward Deployed Engineer. I have been doing the job for three years under a different title: a vendor engineer embedded in a customer's estate, learning their domain, building inside their systems, and leaving tooling behind that their own engineers use.

Read the full case →

A project tracker with a knowledge graph that AI agents keep current themselves. Projects hold tasks, memory notes and links. Claude Code and GitHub Copilot read and write the same store over MCP, so what one agent works out survives into the next session.

Eleven MCP tools, zero runtime dependencies, one local SQLite file, and an audit log that makes every change reversible.

• An invented answer is worse than a missing one. Delphi's knowledge graph extracts deterministically rather than asking a model, because a missing node makes a query return less while a wrong node makes it return something false.
• Undo should not be a second system. Every write records the row as it was, so reversing a change replays stored state instead of guessing at it.
• The tooling you leave behind is the deliverable. The most useful thing I built on my current programme was not a feature, it was the shared config layer that stopped a team of engineers each inventing their own way of using AI.