Andrew Hunter

I’m a systems architect and builder. I build things to work through questions that interest me, and I write about what I find. Lately that has meant transformer inference, GPU performance, and following a question about tokenization back to Claude Shannon’s 1948 paper. The connections between fields are a large part of what keeps me interested.

My route here came through the Navy’s nuclear power program. I learned systems by operating them, then by teaching and supervising the people who operated them. I also learned technical writing through operating policies and procedures. The reader had a plant to run, so an explanation had to help them understand what the system was doing and what a change would cause. I carried those habits into software: work through the behavior, check the explanation against the system, and listen to the person who knows the part you haven’t seen yet.

I publish my active project work at Plainsight Systems, where you’ll find the code, demos, and write-ups as the work develops. It’s where I work through those questions in code I can inspect and measure. Charlotte is an inference engine for the browser, built by hand so I can follow a token through the calculations that produce it. In Serenity, I’m approaching GPU performance through rendering, exploring techniques that also matter to inference: processing data in small pieces, keeping a small amount of running state, and reusing earlier work. Moving light gives me an image to inspect alongside the timings.

Writing and teaching are part of how I work through that understanding. In Let’s Build a Transformer!, I’m building an inference engine in C# with the reader, introducing each operation through a small addition to a running program. The other essays range across architecture, delivery, AI, and transformation. They follow the same interest in why a system behaves as it does, how we know, and what happens when we try to change it.