About us

Rob Nash

I’m a software engineer, and I’ve worked through several shifts in how software gets built. Startups, corporations, agencies, and my own projects, on products for finance, healthcare, and aviation.

The shift happening now is different. This time, the tools can write the code.

What that means for a working engineer is still unsettled. The loudest voices are certain in both directions, the job market hasn’t decided, and I’d rather not wait to be told where I fit. So I’m finding out for myself, the only way I trust. I build real products with AI, ship them, and write up what happens, so whatever I conclude is something I’ve proven rather than something I’ve read.

That standard came from somewhere. I read Biochemistry before I wrote software, and finished with a research masters. That’s a research project ending in a defence, not a taught course. You design the protocol, fix the pass mark before you start, and defend the result to specialists looking for the hole. The subject has nothing to do with software. The habit does, and it’s why I’m careful about what a result proves and what it doesn’t.

With clients and revenue to protect, handing the work to AI is a gamble most can’t take. I can. If an experiment fails, nobody is harmed, I’m out nothing but my own time, and I’ve still learned something a company would rather not learn the expensive way. The freedom to be wrong is what makes the findings worth reading.

Audient is the clearest proof so far. It’s a macOS app that transcribes audio and video and lets you search what was said, on the Mac App Store now. It was built with AI, and it runs AI on-device, including Apple’s own models. There’s an Apple TV app on the desk alongside it, and iPhone and Apple Watch work that’s shelved rather than quietly dropped. Open source comes out of the studio too, and Iris is the first of it public.

Here’s the first thing I’ve found, and it surprised me. The bottleneck was never the typing, and it turned out not to be the reviewing either. I don’t read every line any more. I built a system to do that instead, called Ordova: agents that review the code, enforce the conventions, check the copy, and watch for drift. So the work moved up a level, from doing the job to building the system that does it. The new bottleneck is trust. Every run ends with me deciding whether to believe what came back. That call is still mine, and it’s the part that doesn’t automate.

I’ve proven this works for me. Doing it in the field, where the stakes are higher, is a harder question. It means depending on an agent billed by the token, and the more you trust it the more you spend. That’s a hard line item to defend. It gets easier the day these models run on your own hardware at the same quality, and I’ve tested how far off that is. In July 2026 a local 7B model invented details in three of eight summaries against my reference set, fluently enough that a quick read wouldn’t catch it, and the larger model that might have done better needs more memory than my machine has. So not yet. I still think it lands inside 2026, and I’ve kept the test so I can say either way.

Work history and publications are on the CV, with a PDF if that’s easier.