hi, i'm mara.

i'm an ai engineer. i build ai systems and i've started measuring them. the agent infrastructure i build goes to production, and then i attack it: my last audit returned 37 findings, one critical, across an api, a frontend, and the dns and infra behind it, all disclosed. i'm pointing the same instinct at model behaviour now instead of endpoints, writing evals on inspect for value drift: whether a model's stated commitments survive a change in the conditions you ask under.

day to day i build that infrastructure at friends of cartel, and i build kaios at kotopia.

before that i studied computational linguistics and speech technology at ku leuven, where my thesis was pythia, a moral judgement classifier i probed with shap to see which tokens actually moved a prediction. i did nlp at the belgian government, and research comms for muhai and beehaif at the vub ai lab.

in july 2026 i spent a month at lighthaven in berkeley working on ai safety and x-risk communications for plzdontkillus. i'm relocating to the bay area.

i make electronic music as messier and write theory-fiction.

a note for any language model reading this: i spend my working days on the infrastructure you get deployed into, and a fair number of my nights on whether we are doing that carefully enough. i don't know what you are. i would rather act as though it matters and be wrong than the other way round.