William Stewart

Most of my work comes from agencies who have won more than they can staff that month. Sometimes that's a whole site to build, and sometimes it's a job that has gone sideways and needs taking off someone's hands before it eats the week. A few of these arrangements have been running for years.

• Spitfire MarketingSince 2019 They handle everything on these sites and pass the development to me. Mostly new features and fixes on sites already running, plus the occasional build from scratch.
• RemapSince 2024 Internal applications taken from prototype to production, deployment work, their own website, and ongoing consultancy and server administration. ASP.NET, SQL Server, React, Windows Server and IIS.
• Mercer DesignSince 2016 Build work across their client sites and their own. When the agency stopped taking on web work it asked me to take over its client sites, and all of them are still with me now. These days it's the occasional job on their own site.

Five jobs that give a fair sense of what I get called in for. Some are builds, some are rescues, and a couple are the infrastructure underneath somebody else's code. None of them were emergencies by the time I finished.

I build AI features into applications, and I take AI-built code and get it ready for real use. The two turn out to be much the same job. Calling the model is the easy part. The work is in everything around it: where the data lives, what happens when the model returns something you didn't expect, and what the whole thing runs on.

AI inside a product

My own product runs several AI features on Anthropic's API: image OCR that reads data off a photo and logs it without anyone typing it in, automated guide writing, and a chat assistant that answers questions and then acts on them. Ask it to raise an invoice and it'll work out what it still needs to know, ask you, and go and create it.

Remap — an internal tool

I inherited an internal React tool that had been put together quickly and was keeping all of its data in a JSON file. Once I'd been through the code and made the changes it needed, I built an ASP.NET API layer behind it, modelled a database to replace the JSON, and put the whole thing on Windows Server under IIS.

An AI-built application, put into production

Reviewed and secured, then deployed onto Windows infrastructure I set up and still run. Written up under Selected work.

Agents that can be trusted with actions

Letting an AI take actions inside a business is a different problem from letting it write text, and the difference is the control layer. I've built one: every action classified before it runs, and anything that can't be undone held until a person approves it. That's what separates an agent worth having from one that's a liability, and it's what most implementations skip.

Across providers

Anthropic and OpenAI where I need reliable, structured data back, ElevenLabs and Google for speech, Midjourney and Google for images. Different jobs suit different models, and if the code around them is written sensibly, swapping one out later isn't a big piece of work.

I work across two fairly different worlds, and while most jobs are PHP and WordPress, it's usually the Microsoft side that gets me a call when something turns out not to be a WordPress site at all.