days of mri

mo. nineteen. i build things with ai, and once something clicks, i learn it ridiculously fast.

render=
01

about

the short version
base
london
writes
rust, swift
studied
business, not cs
status
building

most of what i make is tooling for ai agents: the harness they run in, the benchmark that grades them, the limits that keep them in line.

the question i keep coming back to is whether an agent tells the truth about what it did. i don't take "done" for an answer until something has checked it.

i studied business, not computer science. everything technical here i learned by building it.

i care how things feel. a size limit on every binary, a budget for every token, three colours instead of thirty.

little projects here and there.
a couple of big ones.

02

work

big ones first
  1. a overseer an agent harness written in rust. it runs coding agents in your terminal or browser, keeps every session as a log you can rewind, and puts hard limits on what an agent can spend, touch or run. ovrhq.dev
  2. b obench1 an open benchmark for agent harnesses. six harnesses, one model, 27 tasks, the same moment. a do-nothing agent scores zero on every one. open source
  3. c auteur audits my first little business. soon

and the rest

  • woidrag a file, tap a key, drop it on a slice. a wheel under your cursor that converts, compresses and moves files, and never touches the internet.
  • barsmy custom fork of codexbar. how much of your ai limits you have left, and when they reset, sitting in the menu bar.
  • bug huntsi read other people's code for fun. an audit of super productivity turned up a run of timezone and date bugs, with the fixes sent upstream.
03

aims

where this is going
  1. i.

    to build a company, not just projects.

  2. ii.

    to work from wherever i happen to be.

  3. iii.

    to build something great in ai, with people who care about it as much as i do.

if that last one sounds like you, say hi.

04

find me

i reply