Application of AI and LLM in Development and Management (AI column)

Watched this week. speech Alexander Lukyanchenko from the conference AvitoTechConf 2025which I visited in person (But I spent most of my time not looking at the papers, but talking to people I knew and talking about the same topics, but more openly.). If you go back to the report Sasha, he shared the figures about how AI really works in the development processes inside Avito. Sasha leads the development of PaaS within the company and his team is responsible for efficiency. 2000+ Engineers, internal tools, cloud and SDLC.

The main points of the report are similar

⌨️Coding isn't the whole job. Writing code directly takes everything. 20–40% engineer's time. The rest are communication, systems design, review and archaeology. (codebreaking). AI should help here, not just write lines. There are other scenarios, for example. 🗺 Auto-mapping architecture In microservice architecture, it is difficult to know who is responsible for what. Avito used LLM to analyze the code, API and README of all services and decompose them by domain. Result: Automation coincided with manual markup of experts on 80%. Saved ~200 Man-days of handmade architects. ☠️ Postmortem analysis Feed the LLM base from 800+ incidents (postmortems). The model found 22 It's a system pattern of problems that people haven't seen and suggested scenarios for Chaos Engineering. This allowed us to close>1000 potential vulnerabilities. ⚙️ Evolution of tools Guys in the industry are moving from the Copilot phase (supplementation) Agents phase (autonomously). In the top now tools like Cline, Roo Code and agent modes in IDE, which can “walk” on the project and edit files.

What it means for the industry 1. Feelings of productivity are deceptive. Engineers often feel they’ve gotten faster with AI, but metrics say otherwise. (especially the METR study on 16 engineers that I handler). If AI writes a lot of code, which then needs to be debugged for a long time, it is not acceleration, but the generation of tech debt. 2. Greenfield vs Brownfield. AI perfectly boosts the start of new projects (before 30-40%)But on old, complex Legasi projects. ("brownfield") Productivity growth falls to 0–10% and sometimes becomes negative due to rework. 3. Shift of focus. The main value of AI is not in writing code, but in reducing cognitive load. (quick dock search, Sammari endless threads in Slack, Legasi explanation).

P.S. Sasha referred to a Stanford study of 120k engineers. Recently Yegor Denisov-Blanche (Yegor Denisov-Blanch) I told you a new report, “Can you prove AI ROI in Software Engineering?” on this topic on the AI Engineer Cofne and I already have it. handlerThere’s a lot of interesting stuff to watch.

#AI #DevOps #Engineering #Management #Leadership #Software #Architecture #SRE