Alfa-Bank turned a tester’s idea into an AI stub generator for testing teams

An Alfa-Bank tester turned an idea into an AI-powered stub generator for testing teams

An Alfa-Bank tester created an internal AI tool that converts text descriptions into ready-made API stubs for service testing. The project started as an evening experiment within the AI community and eventually grew into an MVP, a separate domain, and testing within the bank.

How the project appeared

The story began in spring 2025, when testing engineer Stas Zaitsev decided to dive deeper into neural networks and explore how they could be applied to everyday QA work. After several experiments, including a Telegram bot for English assessment, he joined an internal AI community around the AlfaGen platform. At one of the meetings, participants were shown a list of useful ideas that teams hadn't gotten around to yet.

Among them was a task well-known to testers: automatically creating API stubs. The idea was practical and very down-to-earth. When the frontend is already ready but the server is still in development, testing often stalls due to the lack of working responses from the service.

Stubs solve this problem, but their preparation requires time, manual configuration, and understanding of infrastructure. In the business requirements for the new agent, a simple scenario was described: the user specifies conditions, the system clarifies details, and provides stub code that can be used immediately in tests without lengthy preparation.

How the generator works

Zaitsev created the first version very quickly in the AlfaGen Prompt Workshop: he set up a scenario that generated stub code based on a description. However, this format proved inconvenient for real work and didn't scale well for colleagues without technical background. Long chat dialogue, manual execution, and the need to understand details made the solution more of an idea demonstration than a tool for everyday QA team work. This is why a second version with a separate interface and automatic execution was needed.

"It was assumed the project would take 3 months, but I didn't know

that and did it all in an evening."

In the second version, the author assembled a full-fledged agent from three parts: WireMock for running stubs, an API layer on Kotlin and Spring AI to connect with the AlfaGen web interface, and a UI that can be used without deep programming knowledge. The user simply describes the needed scenario in the chat window — for example, what responses to return for different userIds, whether a delay is needed, and when to return a 404 error. After that, the system goes through several steps:

• determines the stub type — REST or SOAP
• selects the required HTTP method
• calculates how many separate stubs need to be created for different conditions
• generates code and checks it for errors
• launches the result in WireMock

According to the author, the entire cycle takes 10–15 seconds, after which the stub can already be connected in the working environment. For those who need full control, the interface also has a manual mode where parameters are set separately. This makes the tool useful for both experienced engineers and testers who don't want to spend time on manual assembly and mock deployment. Essentially, the agent removes almost all routine mechanics from the user.

From MVP to launch

The evening prototype quickly grew into a two-week MVP, and then into a full-fledged internal product. After a presentation at the AlfaGen platform demo, other bank teams became interested in the tool, and the developer joined a group already working on QA agents full-time. The project then began the process of reaching industrial quality: they added chat mode, project folders, stub groups with logic, feedback fixes from colleagues, and a more convenient user scenario.

By the end of November, the team closed critical issues and began preparing for a wider rollout. For the solution, they allocated a separate domain, configured access, rewrote the frontend under internal bank libraries, and prepared documentation. Currently, the generator is being tested by focus groups within the bank, and the local version is already being used on real tasks in products for small businesses.

In essence, this is an example of how an internal AI platform turns individual engineering pain points into an applied service for an entire organization.

What this means

Alfa-Bank's story shows that useful AI tools for development can emerge not only from dedicated R&D teams, but from the bottom — from specific pain points of testers. If such projects are quickly brought to interface, documentation, internal standards, and real pilots, they stop being demonstrations of LLM capabilities and become working tools that save hours for entire teams and accelerate the deployment of new services to testing. For corporate AI, this is more important than any beautiful demo scenario.