I build AI agents that do real work

Tyche, the agent that audited the state

Barometrul Digital started from a simple question: how well do public institutions defend their own websites? To answer it I built Tyche, an AI agent that audited 2,487 institutions, from rural city halls to energy operators.

The result left my own audience within hours: 1,277 shares and pickup by Euronews România, ProTV, Economedia and PressOne. In the first week, 41 institutions across 24 counties opened a conversation and started fixing things.

Read the full case study

What an agent actually means here

A useful agent takes an objective, chooses its own steps, uses real tools and stops when it can demonstrate the result. The difference between a demo and something that runs in production is what happens between those steps.

• Deterministic checks. Every conclusion passes reproducible controls. The same inputs produce the same result, every run.
• Traceable sources. The agent works on data you can open yourself, not on what the model remembers.
• Human-in-the-loop where it counts. Before any irreversible action or public claim, a person reviews the output.
• An audit trail. Every step can be reproduced by someone else, in the same order.

How I work

1. We start from a task that hurts. Something the team repeats often, at high effort and with real risk of error.
2. We map the workflow. What comes in, which decisions get made, where verification happens, who signs off.
3. I build the agent with its checks beside it. The deterministic part does the validation work, the model does the judgement.
4. It runs alongside your people. We compare outputs until every difference has an explanation.
5. It goes to production with monitoring. What is not measured cannot be repaired.

What you get

• An agent running on your real data, not on a test set
• The methodology written down, so anyone on the team can explain what the agent does
• The automated checks that keep it honest
• Documentation for the people who will own the system
• Hosting and maintenance, if you would rather not carry that

When it does not make sense

If the workflow is written down nowhere and nobody can say how the decision gets made today, an agent will amplify the existing mess. In that case we start with the workflow map, which is worth having anyway, and only then decide whether automation earns its place.