I build AI tooling that is cheap to run and hard to misuse.

I’m Mark Sulleza, a technical lead with eight years in software.

Most of my day runs through Claude Code and the agents around it. On my own time I build AI tooling for myself, and the hardware and apps around the house. And I like old machines: computers with beige cases, cars with chrome on the dash.

How I work

Read-only until proven otherwise.

Agents that touch production read through read-only database users. The few that write go through the same checks a person’s writes do.

The machine owns the verdict.

An agent saying “done” is a request. Tests, exit codes and evidence decide, and findings come back as supported, contradicted or unverifiable.

Cheap models do the typing.

A strong model plans and checks; cheaper ones write the code it specifies. Every session’s spend is logged, so the saving is a number, not a feeling.

Write the method down.

Skills carry the method, the gotchas and the red herrings already chased, so the next run starts where the last one stopped.

Unknown is an answer.

A confident wrong cause costs more than an honest “the data can’t say”.