I'm David. I build AI systems that do real work.

I'm a hands-on AI systems builder. That means I don't just prompt chatbots — I design and operate environments where multiple AI agents research, write code, control browsers, generate media and trigger automations, with me supervising and steering.

The stuff on this site isn't a curriculum of things I've read about. Every project page describes something that exists, runs (or ran), and produced output — including the ones that are still prototypes. I'd rather show you an honest half-finished system than pretend everything is polished.

I treat AI like a systems engineering problem: inputs, orchestration, tools, outputs, feedback loops. An LLM on its own is a demo. An LLM wired into n8n with MCP tool access, local generation for media, and self-hosted infrastructure is a system.

I build in public with myself: every system here started as a question ("can agents produce a monetisable video end-to-end?", "can a voice agent actually handle a business call?") and got iterated until the answer was yes — or until I learned precisely where the limit is.

• ▸Prototype fast, then harden only what proves valuable
• ▸Self-host wherever it makes sense — cost, privacy, control
• ▸Automate the boring 80%, keep human judgement on the critical 20%
• ▸Document honestly: architecture, trade-offs, current status

Because for the first time, one person can build systems that used to need a team. Research, software, media production, operations — AI collapses all of them into orchestration problems. I find that genuinely exciting, and I've proven it to myself: this portfolio, the videos behind my channels, and the automation running them were all built through the environment described here.

I didn't come to AI from a research lab. I came to it the way most practical builders do — following problems. Building websites led to automating website builds. Running YouTube channels led to automating video production. Needing cheaper, private generation led to local models. One system demanded the next.

That's also why my knowledge is deep in some places and deliberately shallow in others: I optimise for what lets me ship working systems, not for covering every topic evenly.

My YouTube channels are the proving ground for everything here. Four channels run through the AI content pipeline; one is monetised and earns $200–550/month. That revenue matters less as income than as evidence: real viewers, real retention data, real money — produced by systems I built.

Build first, study second. I learn a tool by making it do something real within days, then backfilling theory where it bites. Documentation, source code and primary sources over tutorials. And I keep a low tolerance for "it works on my machine" — if a system can't be reproduced, documented and handed over, it isn't finished.