I build tools that take the friction out of work I do repeatedly. If something costs me an hour twice, it turns into a product. Most of them start as something I needed at 2am and only get released once they survive a few weeks of me actually using them.

Most of my attention right now goes to AI systems — the layer that sits between a model and real work. Orchestration, context handling, tool routing, and knowing where a model will quietly get something wrong so you can build the check that catches it before a user does.

The other half is deliberate and slower: the layer underneath. AI makes it very easy to ship code you cannot debug, so I keep the habit of going one level below what a job strictly needs — hotpath is where that happens, and not one number in it is remembered from someone's blog. I use AI hard, every day. Being able to tell when it is wrong is the entire reason the fundamentals are worth the time — otherwise you are not building anything, you are just moving the bug somewhere you cannot see it.

Faster versions of ordinary things, with the measurement to back it up. Counting newlines seven ways — naive, SWAR, memchr, bytecount, hand-written AVX2 behind a runtime feature check, threaded — measured against what a single core can pull from memory, so it is obvious at which point the code stopped being the bottleneck. Every entry needs a baseline, a named mechanism and error bars, and the results table is generated from criterion output rather than typed by hand. The losses get published too: a table of nothing but wins is a table nobody should believe.

The panel system this profile is built on. Palettes are derived in OKLCH from a single hue and assert their own contrast ratios, so an unreadable theme fails at build time instead of shipping. Superellipse corners, film grain, one motion idea per panel — all of it built inside GitHub's actual rendering envelope, where <style> is escaped to plain text and webfonts never load.

The MCP servers below came out of a stretch of digging into how AI clients actually talk to external systems. Not where my attention is now, but they still run and they still work.

Company lookup against Enhetsregisteret, the Norwegian business register: organisation numbers, roles, subunits, live updates. Ask an AI client who sits on a board and it answers from the register rather than from memory.

Norwegian addresses, place names and elevation from the open APIs of Kartverket, the national mapping authority. Geocoding and elevation data without signing up for anything.

The discord.js v14 boilerplate I got tired of rewriting. Dynamic command and event loading, cooldowns, permission gates, context menus, autocomplete, hot reload.

Private repos. Happy to walk through any of them, or give a lærebedrift read access. vellum is the one with the process on show — every feature lands as a reviewed pull request against a numbered ADR.

aether is the one whose shape I can show without showing the code: