An agentic QA pipeline inside the client’s own repository
Starting point
The client develops almost entirely with AI and questioned whether external quality assurance was still needed at all. We proposed building the quality layer not next to the agents but with them.
What came out of it
The test case catalogue lives as files in the client’s Git. The AI maintains it in the same merge request as the code — no test management tool, no licence. A gate sits in CI that nobody can route around: blocking on the main branch, ahead of deployment. Changing the substance of an already accepted case requires a human sign-off. Repair loops are capped at three cycles, and where a product defect is suspected the agent stops instead of repairing.
Tooling
Playwright · TypeScript · GitLab CI · Claude Code · Coder
Numbers
165catalogue cases, as files in the client’s Git
162Playwright tests generated from them
9.4 minuntil CI says whether the build holds — the whole active suite, one after another
notest management tool — the catalogue lives in the client’s Git
The CI step blocks, it does not merely report: on the main branch, before the deploy. And the AI step has to earn its place — the plain path mechanics without AI reach precision 0.78 and recall 0.88. That is the bar. If the agent does not clear it, it goes.
digital health, Chemnitz · since 01/2026
EndoHealth
Test strategy and a regression suite for Android and iOS — greenfield into the pipeline
Starting point
A Scrum team with a health app on Android and iOS. Until then, testing was done by hand before every release. The brief: test strategy and automation from the ground up.
What came out of it
Test strategy first, then the build: UI automation for both platforms from scratch, running against real devices via BrowserStack, plus the API tests behind it — all wired into the Bitbucket pipeline with notifications into the team. The regression runs on every merge to the main branch and once a day. We started the classic way, by hand; by now we build agentically with Claude Code: agents draft, review and extend the test cases, and the automation itself is improved the same way. Alongside that, test management coaching for the client’s own QA staff.
KfW — German federal development bank, Berlin · 09/2023 – 12/2025
KfW
A test strategy that was workable under banking supervision — and partly implemented by us
Starting point
Around 150 people had been brought together in one programme to build a shared CI/CD component — the pipeline that relevant parts of the whole bank ship through today. On top of that, a deliberately complex architecture, a great many environments, and approval paths that allow no shortcuts under BAIT and MaRisk.
What came out of it
Our brief was the test strategy for that component — one that actually works under the constraints and not just on paper. We implemented part of it ourselves: took over test management, trialled and introduced tooling, trained staff in test automation, and carried a large share of the coordination across the bank.
Regulation
BAIT and MaRisk until they were superseded, DORA from its application date in January 2025. Participation in the Agile Release Train.
Tooling
Robot Framework · Cypress · Playwright · Testcontainers · Azure Test Plans · HP ALM · GitLab
Numbers
150people in the programme, organised along SAFe
3rulebooks the pipeline had to satisfy: BAIT, MaRisk, later DORA
27months in the programme, from 09/2023 to 12/2025
Bosch and Siemens, Munich · 01/2024 – 06/2024
BSH Hausgeräte
Four layers, one test setup: firmware, device, app, voice assistant
Starting point
A new IoT device for the kitchen, together with a mobile app and an Alexa integration.
What came out of it
We tested the device as one system rather than in parts: firmware flashed onto the test unit, swipe sensors, microphones and speakers verified at the embedded level, the app on Android and iOS, the back end behind it and the voice integration on top. We planned, wrote and ran end-to-end tests across app, device and back end, plus the test automation for the mobile side and a test strategy for the build-out.
Web shop, ERP and digital asset management as one programme
Starting point
Fanatec is one of the international leaders in professional sim racing hardware. A custom ERP, a new web shop and a digital asset management system were introduced there at the same time, by three suppliers with more than seventy people who had to deliver together. There was no test or automation practice in the company.
What came out of it
We built test plans, test processes and the automation from the ground up and signed off the external suppliers’ deliveries. The front-end automation ran across web shop, back office, ERP and digital asset management in separate integration and staging environments, plus a load simulation for the Black Friday business.
Tooling
Katalon Studio mit TestOps · Java · Gradle · Postman · TestRail
Numbers
3systems in one chain: shop, ERP, digital asset management
70+people from three suppliers who had to deliver together
