Sixteen years building backends that have to survive launch. Since 2025 — with models inside them.
Below: what I solve, what proves it, and how working with me goes.
Usually it is one of these four.
The model goes where a person already repeats one operation by hand. The success metric is written down before the start. I do not sell a loop that is not there.
FastAPI or Django from the data schema to deploy: cache, indexes, queries, monitoring. So you are not putting out a fire after launch.
Legacy, the real bottleneck, a migration. I get into context, find the actual problem, and bring it to a working state.
Bots, ETL, CRM, external APIs, logging. Python, with support after handoff.
Five years at PVS-Studio. My own systems in production since 2025.
Lead Python Backend / R&D · 2020–2025
Developed pvs-studio.com and ~20 related projects. SmartCache on Redis, 152-FZ web infrastructure compliance, a logging platform with a Python SDK, migrated the site from Python 3.9 to 3.13.
Independent / R&D · since 2025
A public EdTech product: a person sees where to grow and sits a mock interview. FastAPI, PostgreSQL, Next.js. Lives at tech-path-finder.ru.
I take the job when there is no ready spec and no stock solution. I read the system, see where the theory parted ways with production, and fix that. Legacy, migrations, cache, integrations: ordinary work, if you know what you are being paid for.
I plug AI in where a person already runs the same loop and the result can be checked. I do not sell a loop that is not there.
I'll say whether a model fits there or not. I reply within a day. I sign an NDA before the detailed conversation.
I also run AI Builders, an open showcase of dissected LLM systems with a complete textbook underneath it, plus a Telegram channel.
I also run Potapov OS, a Russian-language system for people who sell services.
Short books about the work. Free, PDF.
How to spot the technological shift before the market and collect AI dividends
For people who decide: where AI can already be built into the work, where it is still too early, and how not to buy a pretty imitation of transformation.
Free · PDF · 5 parts · 22 chapters
Open the bookHow to read query plans and make the database do less work
A practical mini-book for anyone who has seen the slow query log: why one index saves your evening, another changes nothing, and a third quietly chokes writes.
Free · PDF · 5 parts · 20 chapters
Open the bookOutside engineering I make melodic heavy metal under my own name. Two albums on streaming: Yandex MusicSpotifyApple Music