_{Leer en español}

I make sure software does what it promises. I build automated checks that use an app the way a person would and raise a flag the moment something breaks, and I wire them into the pipeline so they run on every change. Lately I also check AI: I test what it builds and how it is graded, and I break things on purpose to prove the checks really notice.

Three projects, one per kind of work. Each one says the problem, what I built, how to check it yourself and what was measured. The numbers come from lab runs, not from production systems.

1 · Test automation for a release: restful-booker-quality-suite

• Problem: decide whether an API can ship, and keep known bugs visible instead of hiding them.
• What I built: risk-based Playwright API suite with Zod contracts and latency budgets, run in four CI gates against a pinned Docker copy of the app.
• Check it: docker compose up --build -d && npm ci && npm run quality
• My part: a demo run as a release. I set the risk model (P0: unauthorized changes, booking-lifecycle integrity, authentication) and what blocks a release; I kept the 8 defects visible and wrote down every exclusion. AI assistants wrote the Playwright/Zod code and the report under my direction.
• Measured (lab): 43 checks pass in 2.0 s; 8 confirmed defects kept as expected failures, so a fix turns the suite red until it is acknowledged.

2 · Evaluating an AI agent: agent-eval-jev

• Problem: know when a tool-using agent invents numbers or skips a step, without trusting one LLM judge.
• What I built: code checks first, a calibrated Jev judge second, an LLM judge only for comparison, and a CI gate with a human-review band.
• Check it: python -m agenteval evaluate && python -m agenteval gate
• My part: I decided what to test, the criteria weights and the threshold rules; an AI assistant drafted the 60 reference labels against my rubric and I reviewed them, with no case-by-case log of that review; AI assistants built and ran it under my direction.
• Measured (lab): the evaluator agrees with the reference labels (AI-drafted, reviewed by me without a case-by-case log, no independent annotator) on 93.1% of 29 held-out cases (kappa 0.86); the same gate passes agent v2 and blocks v1.

3 · Scientific QA: predictor-debugger

• Problem: check whether a neural network that predicts a 3D fluid respects physics, without moving the goalposts after seeing the data.
• What I built: a physics debugger with hypotheses and thresholds committed before each round, run on several machines.
• Check it: uv run --locked verify.py --saved mac (minutes, no data download)
• My part: I orchestrated the rounds: three machines and operating systems, pinned versions, publish failures as failures, a reproducible v1.0. The AI proposed hypotheses and thresholds, which I approved before each round, and wrote and ran the code.
• Measured: 49 pre-registered hypotheses over four rounds, 48 pass and one (H14) failed and is published as such; the same fingerprints in six environments.

I lead as an individual contributor: I set the risks, the gates and the rules, and AI agents do much of the execution under them. Three cases, anonymized:

• Stopping production for a false green. In an automated content pipeline I own, an audit showed the image quality gate scored 90 and "passed" while rejecting 10 of 10 shots. I stopped generation, decided the 12 open gates of the fix plan and only resumed when the gate measured again; fix and deploy took about 5 hours. Rule: a gate that says OK without evidence is worse than no gate.
• Release criteria for 10 small client apps. I set the definition (P0 is whatever blocks today or creates legal or real-money risk now; an unrequested feature is never P0; a false claim in the docs is P0) and the rule that nothing ships without smoke and E2E green and a reviewer who is not the fixer. One review round, run by AI agents under that protocol, found 17 P0s and closed 16, with about 125 new tests.
• Correcting AI with evidence. When an AI assistant recomputed my public experiment and called it an independent review, I made it relabel the run as AI-operated under my direction, and published it as a separate release (repro-ia-20261026). Who ran something changes what you can claim.

Senior SDET with 7+ years in test automation, mostly in fintech and banking. I design automation frameworks from zero, test APIs and web apps end to end, and put quality gates into CI.

Lately most of my work is testing what AI builds: making verdicts measurable when outputs are not deterministic, and using AI assistants to design and explore tests faster while a person still signs off.

Ordered from most to least important.

• Golden flows first, then mutate them. One change per mutant; if the suite stays green, that's a finding.
• API first, UI for evidence. Fast, stable checks at the service layer; the browser proves the user journey.
• Deterministic verdicts for non-deterministic systems. Pass/fail where the spec allows it, scores against acceptance criteria where it doesn't, and the same input run 3 times.
• Hard rules live in code, not in prompts. Hooks, CI gates and reviews that can't be forgotten.

Web & mobile: Playwright · Selenium · Cypress · Serenity BDD (Screenplay) · Appium API & performance: Karate DSL · REST Assured · Postman · k6 · JMeter Languages: TypeScript · JavaScript · Java · SQL CI & infra: GitHub Actions · GitLab CI · Jenkins · Docker Practices: BDD (Cucumber/Gherkin) · contract testing · test management with Jira/Xray · AI-assisted testing (Claude Code, Copilot, MCP)

Reach me on LinkedIn. Open to remote contract roles in QA automation and SDET.