Work
What I've taught, built, played with and written.
I've been working hands-on with AI since ChatGPT arrived, and with Claude Code since it became serious. Here's the evidence, including the things I built purely to see if I could.
Selected work
Moving 200 language trainers online, in six months
Linguaid · training
- The problem
- The Covid lockdowns closed classrooms overnight. Language trainers who had only ever taught face to face had to teach well on video, straight away.
- What I did
- With Linguaid, I trained them to teach with Zoom and the other online tools the switch required.
- Outcome
- Around 200 trainers trained in a six-month burst, with the records to show it. The work with Linguaid has continued since, and I'm now a jury member and evaluator for their LangueXpert certification (RS7208).
AI assistants for conferences
Visual Hive · Fractional CTO · 2024–2026
- The problem
- Conference attendees and organisers need fast, accurate answers about programmes, speakers and logistics, and a single general-purpose model is rarely the best choice for every one of those jobs.
- What I did
- I was responsible for app development and AI integration. I tested and benchmarked the major models against real tasks, then built Erleah, a multi-model agentic assistant for events that routes each job to the model best suited to it, using the Claude API from early 2025.
NodeGX
Open source · built with Claude Code
- The problem
- AI makes it easy for a non-coder to build an app, and just as easy to make rookie mistakes along the way. And once the main build is done, they're left depending on the AI for every small change.
- What I did
- Rebuilt a visual app editor from the ground up with Claude Code, over about a hundred documented phases, using the method in my guide. It works as a harness: Claude builds inside it through an MCP server, the structure makes rookie mistakes less likely, and the result is a visual app the owner can learn to tweak and maintain themselves, without relying on AI.
- Outcome
- Live and open source at nodegx.io. The clearest proof I have that the method holds up on a large, long-running project.
Also open source: EventHive, free tools for event professionals that you can configure for your brand and deploy. And before all this, more than thirty client projects with Bubble, Noodl and other low-code tools, either complete applications or joining existing apps as a temporary developer. That's where I learned what breaks when real users arrive.
Built for fun with Claude Code
Experiments, proofs of concept and the occasional rabbit hole. Nobody paid for these. They're here because the best way to know what Claude Code can do is to keep pushing it.