Using AI and LLMs in Software Development and Management (AI)
This week I watched Alexander Lukyanchenko’s talk from AvitoTechConf 2025. I attended the conference in person, though I spent most of my time talking to people I knew rather than watching talks. We discussed much the same topics, just more openly. Back to Sasha’s talk: he shared figures on how AI actually works in Avito’s development processes. Sasha leads PaaS development at the company, and his team is responsible for the effectiveness of 2000+ engineers, internal tools, the cloud and the SDLC.
Here is a rough summary of the main points.
⌨️ Coding is only part of the job Writing code takes just 20–40% of an engineer’s time. The rest goes on communication, system design, reviews and “archaeology”—working through other people’s code. AI should help with that work too, rather than just completing lines of code. Other uses include:
🗺 Automated architecture mapping In a microservices architecture, it can be hard to tell which service is responsible for what. Avito used an LLM to analyse every service’s code, API and README and group the services by domain. The result matched the experts’ manual classification in 80% of cases, saving ~200 person-days of architects’ manual work.
☠️ Postmortem analysis The team fed an LLM a collection of 800+ incident postmortems. It identified 22 recurring patterns of problems that people had missed and suggested chaos engineering scenarios. According to the talk, this helped address >1000 potential vulnerabilities.
⚙️ Tools are evolving The industry is moving from the Copilot phase—code completion—to the Agents phase, where tools carry out tasks autonomously. Tools such as Cline, Roo Code and IDE agent modes are now among the leading options: they can navigate a project and edit files themselves.
What this means for the industry:
1. Feeling productive can be misleading. Engineers often feel faster with AI, while the metrics tell a different story. The METR study of 16 engineers that I covered is a particular example. If AI produces a lot of code that then takes a long time to debug, it is generating technical debt rather than speeding things up.
2. Greenfield vs brownfield. AI can give new projects a substantial boost, with productivity gains of up to 30-40%. But in old, complex legacy projects—“brownfield” development—the gain drops to 0–10% and sometimes turns negative because of rework.
3. The focus is shifting. AI’s main value right now lies in reducing cognitive load, rather than writing code: quickly searching documentation, summarising endless Slack threads and explaining legacy code.
P.S. Sasha referred to Stanford research involving 120k engineers. Yegor Denisov-Blanch recently gave a new talk on this topic, “Can you prove AI ROI in Software Engineering?”, at the AI Engineer conference. I have already covered it. There is plenty of interesting material there, and I recommend watching it.
#AI #DevOps #Engineering #Management #Leadership #Software #Architecture #SRE