Matthew "Smiffy" Smith

My interest in AI is practical, not theoretical. I'm a systems architect. I've been in engineering, working with UN*X systems and databases for nearly four decades. I came to AI recently, firstly through a client requirement, then through a real problem of my own: a need for a cognitive support tool for my AuDHD, then expanding that, to include supporting my partner's severe cPTSD. That systems background turns out to be surprisingly relevant. My current work centres on multi-agent coordination — orchestrating LLM instances across distributed edge infrastructure using purpose-built MCP (Model Context Protocol) tools, including context/state preservation, with a local model handling routine tasks and larger models for complex reasoning, and getting the fiddly work done. I'm particularly interested in how small, well-constrained models can outperform larger ones when properly integrated into purpose-built systems — as can conventional logic. Always assess, choose the right tool for the job. If the tool doesn't exist, build it. I care about AI done well, which means I spend a lot of time noticing AI done badly. Long experience building systems that have to actually work in production gives you a visceral sense for projects that haven't thought about failure modes, data integrity, or whether the problem actually needed an LLM in the first place. I think the industry's biggest risk isn't capability — it's the gap between what people deploy and what they understand. I have explored parallels between RLHF-trained model behaviours and neurodivergent human patterns — there are striking similarities in people-pleasing, pattern rigidity, and context limitations (think: memory issues) that have implications for accessibility. Accessibility for humans and LLMs, both. Building things, breaking things, sharing what's useful.