I build AI products end to end, from the business problem to production. Before writing software full time I spent a decade running retail and wholesale businesses, so I tend to ask what a system is for before asking how to build it.
Most of my recent work lives in private repositories. The links below are the running products — those are the portfolio.
What I'm building
atools.vip — an AI manager for marketplace sellers: Ozon, Wildberries, Yandex Market, Avito. It watches prices, ads, supplies and reviews, catches problems early and proposes fixes, acting only inside the limits the seller sets. Ships a production MCP server (~30 tools) with per-shop OAuth consent and deny-by-default scoping at the dispatch layer. The part I like most: the agent scores its own past recommendations against what actually happened.
ponpon.life — AI personal trainer and nutrition coach. PWA, six languages, iOS and Android betas. The model composes and talks, but every number comes from a deterministic engine, and safety limits live in code rather than prompts. Coach quality is measured on 25 scenarios: took it from 64% to 84%.
sozdai.app — AI infographics for marketplace listings and Avito ads. Live SaaS with paying customers. Layered pipeline: a vision model reads the product, a second model works as art director, a classifier turns user wishes into structural overrides without breaking the hard core. ~1100 tests, CI/CD with zero-downtime deploys and auto-rollback.
camoufox-profile-manager — open-source browser profile manager built on Camoufox. Python/FastAPI, Next.js, Playwright. MIT.
MacCam — offline motion-detecting security camera for macOS. Lives in the menu bar, records HEVC on motion, no cloud. Swift, MIT.
Stack
Python, FastAPI, Celery, PostgreSQL, pgvector, Redis, Next.js, TypeScript, Docker, MCP, Gemini and Claude APIs, Playwright, Swift.
Daily driver: Claude Code (Max plan, daily since early releases); previously Cursor, Windsurf, Codex. In neural networks since 2019, in LLM products since the first public models.