Capgemini Gen AI powered solution: codeai

The Financial Services department in Capgemini Poland has developed a cutting edge solution called codeai that is generating unit tests for Java and Kotlin projects using generative ai. In a classic approach where developers write the tests manually it takes months/years while codeai does it in minutes/hours.

How does it work?

codeai checks out the code and asks ai to create unit tests for each class. Of course the first responses are often wrong but it is very easy to verify them by compiling and running the generated test and checking it against various metrics like static code analysis, mutation testing etc. Only the code that passes all of them is accepted.

What is the output?

Codeai creates a Pull Request for developers to approve or reject so the tool cannot break anything in the code.

The interesting part here is that sometimes ai generates a test that fails on an assert so that kind of tests go to a separate PR that contains potential bugs found by ai!

Is it safe?

Yes. codeai uses a private ai instance that does not connect to the Internet so the mechanism only connects to ai and ai only talks back to the mechanism and the rest is firewalled.

Where does it create the most value?

In legacy projects. You can imagine an old 1 million lines of code java project with test coverage of 5% that nobody wants to touch because something may break. If you run codeai on it it can deliver 10 times higher coverage over the weekend and provide stability plus confidence for the dev team. Of course if you need let’s say 10 people to support these kinds of legacy components you can move 5-6 to non legacy projects if you use codeai. This is huge impact on cost and attrition.

Of course codeai can be applied on any project but if the project is already covered by meaningful collection of tests then its added value would not be as significant as in case of applying it on legacy.

Can I use it?

Yes, if you are a customer of Capgemini or you would like to become one.

Notes

The solution has been also recently described in the “Programista” magazine in Poland: https://programistamag.pl/magazyny/