Barcelona · MCP & AI-agent infrastructure
I build MCP and AI-agent infrastructure and take it from prototype to paying production — auth, quotas, billing, and observability included. Every product below is live: designed, built, and operated end to end by one person.
Before going solo, I grew from frontend engineer to R&D lead at Banki.ru, Russia's largest financial marketplace: a nine-person team, about a hundred endpoints, 2.5 million requests a day, months without downtime — conversion up from 2% to 9%, test coverage from zero to 50%.
Physics and math background from MEPhI. The habit fintech left me with: the real work lives in the distance between a demo and production — so everything in the log ships with auth, metrics, and load tests.
Description & Goals
A living standup catalog: 500+ shows and 300+ comedians in one place.
Semantic search on pgvector: 10K+ transcript chunks searchable by meaning, not keywords.
Built on Next.js and Supabase with n8n data pipelines. Idea to production — solo.
Description & Goals
Seven tools across 11 video platforms: transcripts, subtitles, metadata, chapters, playlists, and search.
Production-grade: Docker images, CI publishing, Redis cache, Prometheus metrics, k6 load tests, optional Whisper fallback.
Powers ComedyCat ingestion in production; works from Claude Code, Cursor, and n8n.
Open-source MCP server that lets AI assistants read transcripts, subtitles, and metadata from 11 video platforms.
Description & Goals
Any MCP server becomes a product: OAuth for Claude, ChatGPT, and Cursor out of the box.
Usage metering and per-tool quotas, with subscription billing on top.
Core: Node.js and Fastify with PostgreSQL.
The next entry. In development now — watch this space.
