AI QA Agents for a Car-Sharing App

At Vloto we are just four developers, shipping a mobile app, a dashboard and an API at high speed on top of several payment systems and detailed charge rules. Testing was always our bottleneck: manual testing covered the obvious screens, but not the edge cases where money or customer trust is at stake.

So I took it upon myself to build an AI-driven QA system. Scripted end-to-end tests run first, including exact-euro money checks. Then AI agents with different personas use the app on their own iOS simulators the way real customers do: a careful customer, one who doesn't pay, one who tries to game the system, a brand-new signup, a family sharing an account, and cheaper sweeper agents that each cover one area.

Every result is checked against the backend and the business rules. A separate verifier agent re-checks every claimed pass and every finding independently, and only confirmed results count. The team rules each finding with one click in a QA dashboard I built, and those rulings become memory the agents read on the next run, so settled items are never reported twice.

A test build reports every screen and API call the agents reach, measured against a catalog of 74 customer journeys on the newest build. The agents only touch staging, test-mode payments and test accounts. In the first two weeks: 52 runs, about 1,270 logged steps, 301 findings flagged and filtered down to 90 confirmed real bugs, 55 of them already fixed, and 64 of 74 journeys verified, at about $6 per agent run.

After a live demo the team decided to invest in dedicated hardware to run it full time.