Projects

The builds behind the research, each made by hand to answer a question.

Each build exists to answer the question on its card, and is built by hand so its choke points stay visible. What each one turns up is written up on the Research page. Code is Apache 2.0 and writing is CC BY 4.0; each repository’s license files govern it.

A small spider in a web of strung lights at night, with words woven into the threads.
updated Oct 2026 · 0★ · 0 forks charlotte What does inference cost when nothing is hidden? A transformer inference engine written by hand in C++, compiled to WebAssembly, with compute in WebGPU and no framework. Running open-weight models in the browser lets me follow the cost of inference through the code, from loading weights to producing a token. Live →Source →
A green parrot speaks into a brass listening horn, with glowing marks becoming handwriting above an open notebook in a lamplit study.
updated Oct 2026 · 1★ · 0 forks dolittle What do we learn from two views of a thinking model? An experiment in comparing a model's generated thinking with natural language descriptions of its internal activations. I read Anthropic's work on natural language autoencoders and wanted to build the machinery myself, so I could inspect both views at the same point in a generation and explore where they agree or differ. Project siteSource →
Glass and brass spheres on a table at night, lit by fireflies that settle into the outline of a small ship.
updated Oct 2026 · 1★ · 0 forks serenity What can rendering teach us about GPU inference? A C++ and Metal path tracer for exploring techniques that also matter to inference: processing work in small pieces, keeping a small amount of state, and reusing earlier work. ReSTIR shares sampled paths across pixels and frames; marbles lit by moving fireflies give me a picture at the end, where people can see what the kernel work makes possible. Project siteSource →
A small electrical transformer rests in a wooden cradle under a brass desk lamp, with a red power lead trailing onto the workbench.
updated Oct 2026 · 0★ · 0 forks cradle Can we build a transformer inference engine we can follow all the way through? The code behind Let's Build a Transformer!, a practical series for .NET developers. We build one growing program in C# with ILGPU, write the GPU kernels ourselves, and introduce the math when the code needs it. Each lesson adds something we can run and check, following Jack Crenshaw's approach in Let's Build a Compiler! Project siteSource →
A relay-bank maze with a small mechanical mouse following a red thread toward the way out.
updated Oct 2026 · 0★ · 0 forks ariadne What would an information-theoretic approach to attention look like? A research program investigating attention through information theory, starting where the field started: Shannon's 1948 paper. Source →
A screenprint of a giant red plant swallowing a stack of memory chips carried by a small figure.
updated Oct 2026 · 0★ · 0 forks seymour Why is a GPU with a petaflop of compute waiting on memory? Take the chip apart, push a token through Llama 3.1 8B, and pull the throttles to see where decode stalls. H100 to MI355X. Live →Source →
updated Oct 2026 · 22★ · 4 forks cpp-perf-guidelines What does a performance engineer know that a model trained on average code does not? The C++ performance technique layer below the ISO C++ Core Guidelines, every example compiled and checked. Source →
updated Sep 2026 · 0★ · 0 forks mcp-servers Can a coding agent use that knowledge while it works? MCP servers in Rust that put the performance corpora in front of coding agents as they write code. Source →

The organization on GitHub: github.com/plainsight-systems →