Stefan Florescu

1. Dec 2022 – present · Remote

   Iron Sheep

   QA Automation & Game QA Engineer (Quality Analyst)
   Client · ZOOT

   A US social & sweepstakes casino platform (formerly Enigma Lake) and its in-house games.

   • Full-spectrum manual QA (scripted, exploratory, A/B) that grew into AI-driven test automation
   • Playwright suite of ~350 tests, Smoke and Regression runs in TestRail
   • Games verified down to the maths: every round re-calculated from the rules
   • Release checks for batches of up to 66 games
   • Built the QA Companion Chrome extension used by the QA team

   Client · AQUA

   NFT marketplace web app. Manual QA across the full range: scripted and exploratory testing, A/B testing, functional and regression testing.

2. May 2018 – Nov 2022 · Bucharest

   Electronic Arts

   Quality Analyst
   FIFA Mobile

   4.5 years on one of the world's biggest mobile games, through its live releases: new-feature, regression and sanity testing, with test cases in TestRail, bugs in Jira and builds from Perforce.

1. Case 01

   Calling a NO-GO

   NO-GOSpec-based testing · edge cases · release decisions

   Problem

   A new slot game's optional "extra play" cost 1.5× the stake, and the spec called it fine.

   What I did

   Turned the spec into ~90 checks and played the edges of the play limits. At the maximum stake the real cost broke the platform limit, and I re-tested in a fresh session before reporting.

   Result

   A clear NO-GO with evidence, so the launch decision was made knowingly.

2. Case 02

   Proving games pay what the rules say

   Game maths · statistics · Node.js

   Problem

   The server calculates every win; I wanted a check that did not depend on the game's own code.

   What I did

   Wrote a verifier per game from the rules document (lines, wilds, cascades, bonus rounds, caps), plus statistical sanity checks (chi-square, z-score, Monte Carlo) on hit rates and RTP.

   Result

   4,000+ real rounds across 6 games, 0 maths mismatches.

3. Case 03

   A QA tool for the whole team

   Chrome extensions · product thinking · Cloudflare Workers

   Problem

   Testers checked balances and banned wording by hand on every game.

   What I did

   Built a Chrome extension (Manifest V3 side panel): a round log, a balance check after each round, a banned-wording scan and a downloadable checklist, shipped with versioned releases.

   Result

   Reached v0.9.3; the QA team used it in their game testing.

• Case 04

  Automation at platform scale

  ~350 tests; Plinko suite 26% faster, no new flakiness.

  Problem & approach

  Problem

  A growing platform and game catalogue needed regression coverage.

  What I did

  Built a Playwright suite (61 specs, 301 catalogued cases) with Smoke and Regression runs, TestRail reporting, a run dashboard and templates per game type.

• Case 05

  Release checks for 66 games at once

  A clear Released / No / Blocked call for every game.

  Problem & approach

  Problem

  A provider delivered a big batch, and each game needed a release decision.

  What I did

  Drove a real Android phone and a desktop browser in parallel: one play per currency, banned wording, currency labels, both layouts and max exposure under the limit.

• Case 06

  Setting the QA process

  Repeatable QA across games; judgement stays human.

  Problem & approach

  Problem

  The QA process lived in people's heads.

  What I did

  Defined how we test (spec to checks, re-test before reporting, clear GO / NO-GO rules) and wrote it down as 10 reusable AI agent playbooks with Claude Code.

1. Built · Sep 2026

   Claw Machine

   BUILTForge (ZOOT's AI game builder) · prompts · reference images

   What it is

   An arcade claw-machine game for the ZOOT platform, credited to me in the archive.

   What I did

   Built it with Forge, ZOOT's AI game builder, by writing the prompts and feeding it reference images and reference games, then shaped it until it was ready to ship.

   Result

   The team turned it into a reusable template: Autumn Claw and Midnight Claw (Oct 2026) are built on it.

   See it

   Claw Machine ↗Autumn Claw ↗Midnight Claw ↗

Tested · a few that stood out

• Cluster cascade slot

  Hellfront

  GO

  Every round re-derived from the rules; hit rate 26.2% against 26.5% predicted.

  What I checked

  Approach

  Turned the game's spec into ~90 checks, wrote an independent verifier for cluster wins, cascades and multipliers, and tested on a real Android phone.

  Outcome

  Released as GO with two minor fixes. A first sample of 101 rounds looked off; 271 rounds showed the game was right.

  See it in the archive ↗
• Coin slot

  Cauldron Grab

  GO

  1,405 real rounds re-checked, 0 maths mismatches.

  What I checked

  Approach

  Verified line wins, Coin Grabs and every Bonus Round frame by frame against the rules, plus balances, languages and layouts.

  Outcome

  GO after two wording fixes, one of them a gambling term in the Chinese translation.

  See it in the archive ↗
• Coin slot

  Coinfall Grab

  FIXED SAME DAY

  A phone-landscape bug found, reported and fixed the same day.

  What I checked

  Approach

  Tested on a real phone held sideways and measured the game window against the layout's limits.

  Outcome

  The play controls were cut off below about 270 px of height; the developers shipped a fixed build the same day.

  See it in the archive ↗
• Line slot

  The Ritual

  NO-GO

  The NO-GO from Case 01: a paid feature broke the play limits.

  What I checked

  Approach

  Played the edges of the play limits in both currencies and re-tested in a fresh session before reporting.

  Outcome

  A clear NO-GO with evidence, so the launch decision was made knowingly.

  See it in the archive ↗

1. 01

   Rules become checks

   Every claim in the rules document maps to a check.

   PASS
2. 02

   Independent verifier

   Re-derives every round from the rules, not the game code.

   PASSSee output ↓
3. 03

   Exact model + data

   All 27,000 outcomes give the exact RTP; 5,000 rounds are tested against it.

   PASSSee output ↓
4. 04

   Playwright suite

   16 tests on desktop and phone; a planted bug fails exactly one.

   PASS
5. 05

   QA report

   Verdict first, then evidence, then every check.

   PASSWATCH 1
6. 06

   Play it

   The game runs in the browser; a seed makes rounds replayable.

   PASS

• assert 01

  Re-test before reporting

  A finding goes out only after it reproduces in a separate run.

• assert 02

  Small samples lie

  18% hit rate over 100 rounds was 26.2% over 271 (26.5% expected).

• assert 03

  Measure, then change

  Speed work starts with a baseline with retries off and ends with three clean runs.

• assert 04

  Verdict first

  Reports open with the decision and the evidence.

Game QA

Live-service mobile games, RTP and paytable checks, statistical sanity checks, provider parity, sweepstakes release checks

Testing

Manual, scripted and exploratory testing, A/B testing, test automation, regression and sanity, test case design, spec-based testing, release testing, risk analysis

Tools

Playwright, Postman, TestRail, Jira, Confluence, Perforce, Chrome DevTools / CDP, Android ADB, Git / GitHub

Code

JavaScript, TypeScript, Node.js, HTML/CSS, Chrome Extensions (MV3)

Infra

Cloudflare Workers (extension release site), AWS ECS + Pulumi (helped deploy game servers)

AI

AI-driven test automation, AI agent playbooks (Claude Code), AI-assisted test design, verification and reporting