Open-source tooling, the platform work that paid for the lessons, and a community that turned into a company.
Turns any repository into a governed AI workspace: quality gates, security scanning and risk management run on every commit.
An agent will happily write code that ships a secret. This puts the same rules in front of Claude Code, Copilot, Cursor and Codex, so the gate decides what lands rather than the model.
A cross-platform statusline for Claude Code showing model, context, rate limits and effort.
It puts model, context and rate limit in the terminal you are already looking at, so you notice you are running out before it stops you.
The same statusline idea, ported to the GitHub Copilot CLI.
Released out of the Arcasiles community after people asked for the Copilot version.
An MCP server that reads CI/CD pipelines on Azure DevOps and GitHub, applies security policy and posts compliance findings back onto the pull request.
Pipeline review is the least loved job on any platform team. This one reads every pipeline and never gets bored.
A 45-minute hands-on workshop that takes a room from install to a merged pull request.
Written for the session I ran at Celonis in May 2026, then left in the open so anyone can run it.
A C# client for the Meetup API, written during the .NET years.
Still forked and still opened, ten years on. It is here because the .NET years happened and I am not going to pretend otherwise.
Standards, API governance and company-wide CI/CD, run as an internal product rather than a policy document.
Governance only works when it removes friction. If a team has to choose between the golden path and shipping, the golden path lost.
A cultural-tech company built on three lines: tech, events and community, with the music side inside events.
It started as people meeting up in the Vega Baja. The company came later, in November 2024, once the community had proved it would keep showing up.
These are private, so I can describe what they do but not name them.
A kernel-isolated, auditable memory layer for AI agents working inside regulated, EU-resident organisations.
Agents forget, or they remember too much in the wrong tenant. Both are compliance problems before they are engineering ones.
The thing that runs a company staffed largely by AI agents: approvals, kill switches, and the real cost per task.
It is also the honest source material for the talk about running twenty agents and surviving.
ai-engineering is MIT and installs in a minute. The rest is a conversation.