Filipe Pawlik Leite
I make AI agents stop lying.
I build the harness around AI agents — the gates, evaluation and orchestration that decide whether an agent's output is good enough to ship. Then I sit with the customer and stay until the thing works.
A gate that
refuses.
Anything an agent writes has to pass typecheck, a real test suite and — for web — a build that produces a servable artifact. Nothing ships otherwise. Flip the switch and break the code. The gate below is the same logic, running here in your browser.
Press run. The gate does not know which variant you picked.
Proven honest with shadow tests: known-bad code is rejected. When the gate fails, the system feeds the error back to the agent and retries with backoff, up to a hard cap.
Every number here
was counted.
If a figure on this page has no source, it should not be on this page. Small samples are shown with their sample size.
Routing by hit rate,
not by vibe.
Measured accuracy per agent decides who gets the next task. Pick a job. The router weighs hit rate against cost and picks — the same way it does in production.
Hospitals, blood banks,
governments, Chevrolet.
Systems that run
without me.
For every layer I ask the same three questions: what it costs, how good it actually is, and whether it holds up as things grow. Then I measure. A candidate is not a decision, and a study is not a decision. The house standard is what survived measurement.
Let's find your
real problem.
filipe@netcks.com
Bombinhas, Santa Catarina, Brazil. UTC−3 — full overlap with US business hours. English, Spanish and Portuguese, all fluent, all working languages.