About

I work on AI agents, and I care about both ends of it: how a model gets called, and how a service stays alive when nobody is watching. Demos don’t satisfy me — if something can only run for ten minutes, it isn’t finished.

Most of my time goes to Autarkia: a self-hosted agent runtime on one machine, Python with an optional C++ scheduler, local GGUF mixed with cloud APIs. 94 tools across 26 domains are assigned per role as resident, on-demand, or disabled; memory uses weighting, vectors, and nightly distillation; proactive messages are bounded by quiet hours and a daily cap.

What eats the most time is boundaries: one source of truth per fact, every config field must be able to answer “who reads it”, and write operations have to be reversible. The self-healing path grew out of the same concern — detect anomalies in the logs, run root-cause analysis, let the patch pass a sandbox, then wait for a human to approve it.

It is built for one person: no multi-tenancy, no SaaS, and the data stays on the machine. The iOS client is still an early implementation, with far fewer features than Android.