Not an Experiment, but a Strategy: A Path to System-Based AI in SDLC (AI column)

There's a tape. speech Nikolai Bushkova, a colleague of mine, who leads the engineering productivity group at the T-Bank R&D Center. By the way, this R&D center was established a year and a half ago, investments amounted to more 500 million rubles. The center is engaged in research in the field of artificial intelligence, databases, information security and engineering productivity. It cooperates with leading Russian universities, including MIPT, Skoltech and NSU. As for the key ideas of the report, they are as follows:

1. Transition from Experiment to System Approach The main concept of the report is that AI at T-Bank is not a hype or experiment, but part of the engineering strategy and everyday working tool. More 5000 Experts take tips from internal assistants per week, which indicates the mass adoption of the technology. 2. Industry review and research Nikolai cites research by Anthropic showing that a third of all cases of using their chatbot account for software engineering, despite the fact that programmers make up a much smaller share of the total number of users. The ability of models to close tasks completely turnkey is already quite high. 3. Systematic approach to use cases The team broke the software development life cycle (SDLC) It has established internal working groups for each stage. The priority scenarios were: Autocompletion of code and chat in IDE

• Unit test generation
• Automatic code review
• Advanced code search.
• Assistants outside the IDE 4. LLM platform The company has its own LLM-platform, which is Own models of T-Lite and T-Pro (Not yet actively used for code tasks) By the way, recently there were announcements of new versions. Serving open weights models like Qwen, DeepSeek Loyal information security department allowing limited access to external models 5. Development plans Development from assistant to agent - work is underway to create more autonomous solutions that can completely close tasks There are pilots with external solutions - testing Aider and other agent systems 6. Methodological principles Nikolai emphasizes the importance of measuring not only adoption metrics, but also user satisfaction. Do what hurts, not what we can do technically. When implementing, it is important to pay attention to quality, so as not to increase technical debt.

In general, the report of Nicholas demonstrates a mature approach to the introduction of AI in the development of software

#AI #Engineering #Software #Leadership #Metrics #Agents