Notes on shipping, design, AI tools, systems, and the occasional personal essay.
What I learned by starting a tiny router from random weights, training it on 100,000 generated prompts, and checking it against real model behaviour.
How I built a 14M-parameter router from empirical model results, exported it to MLX, and learned that the labels matter more than the classifier.
How I moved prompts out of the codebase and gave them variables, versions, API keys, weighted tests, and metrics.
How one backend cleans up fourteen mountain data providers for Traverse, Slab, Outpost, and the next app.
How I turn spoken instructions into local coding tasks and keep the terminal readable while agents work.
What I learned while testing model ladders, confidence bounds, live traffic, and the cost of each passing result.
How I turned glaciology papers, terrain data, weather hindcasts, and validation checks into a crevasse-hazard model.
A quick sprint through the software and hardware I lean on every day, split into tools for thinking, doing, and keeping momentum on the go.
A short piece on what leadership still owns when building gets cheaper and execution is no longer the main constraint.
A compact note on finding signal quickly, tightening decisions, and moving without turning speed into noise.
A take on how AI shifts the cost of articulation and why the old convince-someone-else loop is breaking down.
A short personal essay with lessons, observations, and a few scars collected by twenty-five.
An older piece on using component-driven frontend code to make design systems more consistent and easier to evolve.