- 🎯 AQA Lead / Senior SDET (TypeScript / JavaScript), Warsaw. 8+ years in software quality, 5+ of them in hands-on test automation. I design test systems that let teams ship without fear: test architecture, CI/CD quality gates, mobile device farms, test observability.
- 👥 Today I lead a QA team that owns the quality of a core part of a B2B SaaS product: manual + automated test strategy, a web & mobile framework built from scratch (Playwright, WebdriverIO, Appium), an in-house iOS/Android device farm, and Claude Code agentic workflows that give the whole team one entry point for AI-assisted testing.
- 🌍 Domains: B2B SaaS and fintech — on web, mobile web, native iOS and Android (React / React Native apps under test).
- 🚀 I'm building two products of my own, end to end: Forma — a financial operating system for freelancers, and Zagaday — AI-personalized party games.
- 🚧 In the open, I'm building Flakemetry (test runs as traces) and Sentra (AI triage for CI failures) — and pushing AI-assisted quality engineering well past autocomplete.
- 🤖 My open-source repos have a maintainer of their own: kogut, an agent built on Claude Code that reviews Dependabot PRs, diagnoses red CI and audits repo hygiene — policy in code, judgement in the model.
| The problem | How I solve it | Proof in the open |
|---|---|---|
| "Our suite is flaky and nobody trusts red anymore." | Treat every test run as a trace, not a pass/fail report. Measure what varied between passing and failing runs, make flakiness explainable, and eliminate it with data instead of re-runs. | Flakemetry |
| "Engineers burn hours triaging CI failures." | Put an AI triage layer in the pipeline: classify every failure on two orthogonal axes, hypothesise the root cause, post one actionable PR comment — and hold the model to a hand-labelled golden dataset and a non-LLM baseline. | Sentra |
| "Feedback takes hours — every branch queues for one shared staging env." | Give every pull request its own ephemeral environment: isolated Kubernetes namespace, suite sharded across parallel Indexed Jobs, one aggregated report, and a teardown the pipeline itself proves. | Ephemeral K8s test envs |
| "Our E2E suite collapsed as soon as a second person touched it." | Architecture before scripts: page objects, fixtures and selectors as separate layers, suites split by trigger, a global health check that fails fast with a reason. Built for a team to extend, not just to run. | Playwright E2E framework |
| "API tests are green, but integrations still break in production." | Test contracts, not status codes: schema validation on every response — including across services — consumer-driven contracts with Pact, property-based tests that generate the edge cases nobody hard-codes, and k6 performance budgets. | Microservices API testing |
| "Appium or native? Everyone argues, nobody measures." | Test the same app three ways — Appium, XCUITest, Espresso — behind one CI pipeline with a combined report, so the trade-off is demonstrated on evidence, not opinion. | Mobile automation, 3 ways |
| "Dependabot PRs pile up — and green CI doesn't prove an upgrade is safe." | Split the work: policy in code, judgement in the model. Tested scripts own the merge gate; the AI only reads release notes against the code actually in use and explains red workflows. It can hold a PR back, never merge past a blocker, and I approve every merge. | kogut |
| "We have no automation at all. Where do we even start?" | Build the framework from scratch, wire it into CI/CD from day one, and mentor the team until the system is theirs — I optimise for what keeps working after I leave the room. | every repo here started as an empty folder |
- Start from the problem, not the tool. Frameworks, runners and dashboards are consequences. The failure mode you are fighting comes first; the stack follows.
- Tests are production code. Same architecture standards, same code review, same refactoring discipline. A suite nobody can extend is technical debt with a green checkmark.
- CI from day one. A test that doesn't run on every change doesn't exist. Quality gates, sharding and fail-fast health checks are part of the framework, not an afterthought.
- Make quality observable. Reports, traces, metrics. A red build must mean something — flakiness is a signal to measure and eliminate, not noise to re-run.
- AI with evals, not vibes. I use AI for test generation, failure triage and root-cause analysis — and hold it to golden datasets and baselines like any other system in the pipeline.
- Prove the boring parts. Teardown is tested. Reports are published. Docs let the next engineer onboard without me. Done means demonstrated.
Two products, both mine from A to Z — product thinking, code, infrastructure and quality:
🏦 Forma · @forma-finance
The financial operating system for freelancers, creators and the self-employed. Self-employed people don't have a financial system — they have fragments: a spreadsheet, a government portal, an accountant who answers in three days, and their own memory. Forma is the calm, guided layer on top, answering three questions on every screen: am I okay? what needs attention? what happens next? Mobile-first (React Native, iOS + Android), built on a deterministic, versioned, source-backed rule engine — Poland first, starting with Ukrainians running a JDG.
Built from A to Z: the product itself, the rule engine, the React Native app, the website — and, naturally, the quality system around every release.
Stage: 🟡 pre-release — the mobile platform and the website are in active development · Ukrainian, Polish and English from day one.
🎲 Zagaday · @zagaday
Party word games, personalized by AI. An offline-first party word game (explain-the-word, team vs team) with a twist no clone has: type "Olia's bachelorette party" or "IT company offsite" — and AI generates a word deck that's actually about your crowd. Generated once, stored locally, playable offline forever. Ukrainian-first 🇺🇦, decks are never translated — each language gets its own culturally-native content. And no ads mid-round, ever.
Stage: 🚧 building the iOS MVP — moving along a public six-stage roadmap:
| Piece | Status |
|---|---|
| Product spec, design system, content guidelines | ✅ done — the source of truth |
| iOS app — Swift 6 · SwiftUI · SwiftData · StoreKit 2 | 🚧 in progress → TestFlight, then App Store |
| AI deck generation + Pro (Python/FastAPI backend) | 📋 next — the differentiator |
| Polish & English decks | 📋 planned |
| Android app — Kotlin · Jetpack Compose | 📋 planned — starts after iOS ships |
| Online mode | 🔮 only if traction proves it |
| Project | What it is | Stack | Status |
|---|---|---|---|
| Flakemetry | OpenTelemetry-native test-intelligence platform — test observability, explainable flaky-test detection, AI root-cause analysis; Playwright, Vitest and Jest reporters published to npm, plus a pytest plugin | TypeScript · OpenTelemetry · Fastify · PostgreSQL · Next.js · Helm | ▓▓▓▓▓▓▓▓░░ ~80% · 🟢 active — usable today (M0–M4 & M7 shipped; scale & launch next) |
| Sentra — AI flaky-test triage | AI triage layer for CI failures — two-axis classification, root-cause & fix-suggestion agents, one PR comment; evaluated on a hand-labelled golden dataset vs a non-LLM baseline | TypeScript · Claude API · Zod · Vitest · Playwright · GitHub Actions | ▓▓▓▓▓▓░░░░ ~60% · 🟢 active — full pipeline runs in replay mode (baseline scoring, ablations & v1.0 next) |
| kogut — agentic maintainer | Maintenance agent for the six repos in this table — reviews Dependabot PRs (incl. breaking changes behind green CI), diagnoses red workflows, audits hygiene, opens fix PRs under its own GitHub App, reports per repo to Discord | Claude Code · Bash · jq · GitHub App · Discord | ▓▓▓▓▓▓▓▓▓▓ 14/14 roadmap items · 🟢 live — maintaining this portfolio |
| MoonTrack | Time tracking and time-off management for teams — approvals, leave balances with accrual, team calendar, audit log, isolated multi-tenant organisations — built together with its test suite (Playwright E2E + API specs, pytest) in one monorepo | Python · FastAPI · SQLAlchemy · Next.js 16 · TypeScript · Playwright · pytest · Docker | ▓▓▓▓▓░░░░░ ~50% · 🟢 active — core product runs locally (deployment & production readiness next) |
| Ephemeral K8s Test Envs | Per-pull-request ephemeral environments on Kubernetes — isolated namespaces, 113 API tests sharded across 4 Indexed-Job pods (~3.2× faster), aggregated Allure report, teardown the pipeline proves | TypeScript · Kubernetes · Helm · kind · Docker · Allure | ▓▓▓▓▓▓▓▓▓▓ 100% · ✅ complete — reference build |
| Playwright E-Commerce Framework | Layered E2E framework built for a team to extend — POM, fixtures, selectors, tagged critical / smoke / regression suites, Chromium · Firefox · WebKit CI matrix, Flakemetry reporter, live test-health dashboard | TypeScript · Playwright · GitHub Actions · Allure | ▓▓▓▓▓▓▓▓▓▓ 100% · ✅ complete — maintained |
| Restful Booker API Testing | Production-grade API automation for a Spring Boot microservices platform — nine test layers: Zod contracts incl. cross-service, Pact consumer contracts, property-based, OWASP-oriented security (ZAP), k6 performance budgets | TypeScript · Vitest · Axios · Zod · Pact · k6 · Docker | ▓▓▓▓▓▓▓▓▓▓ 100% · ✅ complete |
| Mobile Task Manager Automation | One React Native app tested three ways — Appium vs XCUITest vs Espresso — behind one CI pipeline with a combined live report | TypeScript · Appium · XCUITest · Espresso · WebdriverIO | ▓▓▓▓▓▓▓▓▓▓ 100% · ✅ complete |
| Portfolio website | Personal site & professional portfolio — Next.js, E2E-tested with Playwright (incl. accessibility checks), shipped via GitHub Actions | Next.js · TypeScript · Tailwind · Playwright | ▓▓▓▓▓▓▓▓▓░ ~90% · 🚀 launching soon |
kogut is an agentic maintainer built on Claude Code, and it looks after six of the repositories in the table above. It reviews Dependabot PRs, diagnoses failing workflows, audits repo hygiene, tracks every finding as a GitHub issue, opens fix PRs and reports per repository to Discord. It acts as its own GitHub App, kogut-agent[bot], so its work and mine stay apart in the history.
Why it exists. The costly failures in a portfolio like this are the quiet ones: a scheduled workflow GitHub switched off, a nightly report that stays red for a week, a major bump whose CI is green only because nothing exercises the code it breaks. A bot that merges whatever is green misses all three.
How it's built: policy in code, judgement in the model.
| Who | Decides | How |
|---|---|---|
| ⚙️ Code | Everything that needs no judgement: the merge gate, failure streaks, flaky and disabled workflows, hygiene | Bash and jq with no model involved, covered by policy tests. The model can hold a PR back, but it can't merge past a blocker. |
| 🧠 Model | Only what code can't: does this major break our code? Why is this workflow red, and what's the smallest fix? | Claude Code reads release notes against real usage, and failed logs against the default branch. |
| 👤 Me | The final call | Every merge goes through a permission prompt. I review and merge every fix PR; kogut never merges its own. |
Built for hostile input. kogut reads text written by strangers (release notes, CI logs), so the design assumes that text is an attack: a least-privilege command allowlist, a GitHub App that can't bypass any ruleset, and updates that never run on my machine.
LinkedIn — linkedin.com/in/andrii-kohut · Email — a.kogut01@gmail.com · Portfolio — launching soon
Always up for a conversation about test architecture, flaky tests, mobile automation, or bringing AI into quality engineering.




