I build tools to understand, evaluate, and improve AI agents.

I’m Luka, an AI software engineer at n8n in Belgrade. These are my independent projects.

A few things I’m building.

Each starts with a question I want to understand better.

Rubrist

Who checks the judge?

I care as much about the evaluator as the answer it scores. Rubrist lets you build a focused check, inspect its mistakes, and compare it with independent human judgments.

Explore Rubrist

A robot and a human comparing different verdicts on the same work.

Overclock

Give the next session a head start.

I want useful lessons to carry over to the next conversation. Overclock gives coding agents reusable workflows for memory, debugging, review, and everyday work.

Explore Overclock

An AI robot powering up with a rising energy aura and levitating stones.

Try your first agent evaluation

Walk through a simple example with your coding agent.

Ironside

The answer is only part of the story.

When an agent gives a surprising answer, I want to retrace how it got there. Ironside keeps inputs, model responses, and tool calls together so you can inspect what happened.

Explore Ironside

A robot unspooling a recorded sequence of its earlier actions.

Where to start

Three of these tools form one path: record what your AI did, turn the failures into checks, then decide whether a change can ship. If you have traces already, start with Rubrist. If you install third-party skills, start with Casefile. Each command below installs the tool's plugin in Claude Code; every project page also shows the plain npm or Docker path.

Ironside
Records what your AI did: inputs, model responses, and tool calls.
Pre-release · Self-hosted/plugin marketplace add luka-zivkovic/ironside
Rubrist
Turns failures into evaluators you can check against human judgment.
Beta · Self-hosted/plugin marketplace add luka-zivkovic/rubrist
Dailies
Decides whether a change meets your release rules.
Pre-1.0 · CLI/plugin marketplace add luka-zivkovic/dailies

On their own

Casefile
Inspect a skill before installing it, and notice later changes.
Pre-1.0 · CLI/plugin marketplace add luka-zivkovic/casefile
Overclock
Skills for coding agents: memory, debugging, review, and evaluation.
Public toolkit/plugin marketplace add luka-zivkovic/overclock

Latest writing.

Guides to the tools, and working notes on ideas I’m evaluating.

Essay

Jev and Claude as judges: 1,459 labelled cases

Cheap judgments, agent traces, and when sending uncertain cases to a stronger model pays off.

All writing →

There’s more to explore.

Meet Casefile and Dailies, and wander through the whole collection.

Explore all projects