AI builds.I verify.

Working software, adversarial tests and public evidence from an early-career builder who directs AI without trusting it blindly.

Choose your proof Open auditable CV

Kuba's auditable CV showing four of four live checks passed — **Auditable CV** Live API checks. No fallback numbers.

Three things I can prove.

A working product, a cryptographic verifier and a review of the bugs that survived happy-path testing.

Zustand DevTools

A Chrome DevTools extension for state inspection, path-level diffs, time travel and shareable trace sessions.

Product judgment / TypeScript / testing / release discipline

DKIM signature verifier recovering a real RSA padding block

Break the proof

A real DKIM signature, two implementations and no cryptography library.

Python / JavaScript / security reasoning

Three silent bugs

The failures that looked successful, and the review methods that exposed them.

AI oversight / debugging / negative testing

The useful skill is judgment.

Specify

Turn a fuzzy goal into constraints, failure conditions and a definition of done.

Direct

Give AI bounded work, the right context and acceptance criteria it cannot negotiate away.

Attack

Test seams, negative cases and assumptions owned by browsers, frameworks and external APIs.

Ship

Release only what I can explain, reproduce and support when the happy path disappears.

257 tests, all green.These three shipped anyway.

Every automated test I had passed. That is the uncomfortable part — none of these defects announced itself, and none of them was a failure the suite could have reported. A green suite is evidence that the cases you thought of still work. It is not evidence that the code is right.

Point at any of the three and the suite goes quiet.

Automated test results 257 passing / 0 failing

A field of 257 marks, one per automated test across three projects — 96 in Zustand DevTools, 130 in DAEMON, 31 in Found Local. Every one passes. The three defects listed below are not in this field, because no test failed.

· one passing test 96 · 130 · 31

01
A tree walk that counted siblings as depth

Any list over forty items silently lost its tail. Fixtures were small, so every test passed at the only scale they measured.
walk(fiber.sibling, depth + 1, out); all 257 tests still passed
02
A shared global assigned instead of extended

It destroyed React DevTools, or was destroyed by it, depending on load order. Nothing was wrong in isolation, and isolation is how it was tested.
window.__REACT_DEVTOOLS_GLOBAL_HOOK__ = { /* mine */ }; all 257 tests still passed
03
A licence check that validated instead of activating

Both endpoints return success for a real key, but only one registers an instance — so the activation limit was never enforced.
POST /v1/licenses/validate all 257 tests still passed

All three were found by a review pass over finished, released code — not by the suite. Read how each one got through →

What you would be hiring.

Not a senior developer. An early-career software builder who gets unusual leverage from AI and treats verification as part of the job.

Best fit: AI-native product teams, developer tooling, internal automation and QA-minded engineering.
Strongest value: Turning unclear goals into tested working systems quickly.
Working style: High ownership, fast feedback, honest limits and evidence before confidence.

See whether I fit your team.

Dublin-based and open to software engineering roles where AI leverage and verification both matter.

Open LinkedIn Email Kuba