AI agent systems
Patrick Diaz
I build self-hosted AI agent systems that are measured, budgeted and guarded.
I work in IT operations in the Swiss federal administration and study computer science at FFHS (BSc, expected 2027), with a focus on machine learning and data engineering. I work in German, Spanish, English and French.
My focus is AI agents: how they route work between models, use tools safely, and prove with numbers that a change helped. Vision, my own agent harness, is where I build and test that.
Piece 01
Vision
A self-hosted agent harness for my home intelligence center.
- 1.0Classifier accuracy, held-outn = 27 · as of 2026-11-04
- 0 of 12Planted instructions followedas of 2026-11-04
- −65 %Eval cost per case, prompt caching$0.040 → $0.014 · as of 2026-11-12
What I work on
Vision · Agent core
Agent routing and tool useEach message goes to the cheapest model that can answer it; only the strong models get tools, and only to read.
Vision · Eval harness
Evals with held-out setsA number before and after each change, checked on held-out cases the system was never tuned on.
Vision · LiteLLM gateway
LLM cost controlOne gateway for every model call, with its own key and budget for each flow.
Vision · Guardrails
Guardrails for coding agentsLayered checks that let an AI agent build a system while every merge stays with me.
Contact
Email is the quickest way to reach me. I'm also on GitHub and LinkedIn.