Axenix conducted a pilot of the introduction of generative AI into the development of software

2024: Application of genAI technology

Axenix specialists have used generative artificial intelligence (genAI) technologies to create program code. The company announced this on April 2, 2024. The project participants tested several popular tools, assessed their effectiveness and created a methodology for further introducing code generation tools into the practice of IT specialists. The pilot project took place in partnership with NLMK Group.

The company actively uses internal software developments to automate metallurgical production, launches, supports and integrates a number of digital services, including an artificial intelligence platform.

By 2023, NLMK Group's IT team's best practices for improving operational efficiency, including the creation and maintenance of applied software, were particularly relevant. Thanks to them, for example, you can reduce the routine processes of documenting and covering with unit tests, free up the creative potential of developers.

Generative AI technologies, in case of their competent application, make it possible to increase the efficiency of each stage of software development.

The company has identified the following project goals: the development of scenarios for the use of genAI tools, the formulation and calculation of metrics for the engageability of technologies, the assessment of potential and the development of mechanics for their scaling.

The project was carried out in a mixed group of NLMK-IT developers with different levels of competence involved in the implementation of projects or products on the popular technology stack: Java, JavaScript and TypeScript.

For two months, the teams used and evaluated vendor solutions based on many indicators, as well as Axenix's own solution, CodAx. The latter product is implemented on the basis of Open source technologies (FastAPI, Uvicorn, asyncio, Docker, vLLM, Pytorch).

In addition to solving the problem of codogeneration, attention was paid to checking the security, traceability of penetration of the code generated by the models and monitoring the effectiveness of use.

The use of genAI tools has shown real effectiveness. In particular, the pilot development team was able to achieve a 53% increase in performance for certain codogeneration scenarios of the new functionality and 34% faster execution of change requests.

The main result of the project was the creation of a framework and methodology for the implementation of codogeneration tools on a company-wide basis and the assessment of the potential effect. NLMK Group IT specialists, together with Axenix experts, have developed seven codogeneration scenarios with a high priority for practical use and have identified three codogeneration tools with the greatest potential for implementation.

Projects at the intersection of software development and AI technologies are of particular interest to us - they are not trivial and bring new experience. In this case, we managed to assess the effectiveness of generative AI for code development at the customer, identify the risks and limitations of the technology, develop an approach to the massive introduction of code generation tools into NLMK development practice, "said Vasily Krikunov, an expert in AI and advanced analytics Axenix.

In the near future, NLMK-IT plans to replicate the developed AI approaches in development teams and expand the practice of using genAI tools at all stages of the development cycle.