Twenty years of production engineering
I started writing production code in 2005 — well before AI agents, well before cloud-native, well before half the languages I now use existed in their current form. Over those years I've shipped real-time multiplayer games on Node.js + WebSockets, a browser strategy game with around a thousand active players on a LAMP stack, a crypto trading bot in Python, and twenty-five-plus client web projects from 1998 onward.
That long tail of production experience is the lens I bring to every modern problem. Architectural taste isn't a parameter you can crank up; it's a habit built from years of fixing the consequences of bad decisions.
Where I am now
My current focus is production AI systems — from single LLM-in-the-loop apps up through multi-agent architectures, with the data layer, evaluations, and operational discipline around them. The mix shifts with what each engagement actually needs; the through-line is "design what should be built, then ship it well."
Since March 2025 I've been leading the AI architecture for a US-based biotech client building autonomous AI systems for scientific research. My charter: own the architecture end-to-end (orchestration patterns, tool-use design, memory and state management, security boundaries, performance trade-offs) and lead the implementation that turns architecture into working software.
I built a specification-driven AI development methodology that turns a written spec into a working system rapidly and reproducibly. I built the evaluation framework — public benchmarks plus custom internal scenario suites the system never sees during development. I built custom MCP servers (Stdio, SSE, Streamable HTTP) bridging LLMs to proprietary tools and secure execution.
How I actually work day-to-day
I don't write code by hand anymore. I direct coding agents — Claude Code on Opus — as the implementation layer, while I own architecture, methodology, and quality. This site is itself an example: the entire monorepo, the FastAPI backend with clean-architecture layers and Protocol-based DI, the Next.js frontend with i18n routing and Tailwind v4 design tokens, the Google OAuth flow with signed state and opaque session cookies, the multi-stage Docker builds, the Alembic migrations, the Railway deploys — all of it was directed, not typed.
What does that mean in practice? Twenty years of systems judgment go into designing what should be built. Daily agent practice gives the velocity to actually ship it. Spec quality becomes the new bottleneck; the discipline transfers from "write correct code" to "write correct specifications and review correct code."
Earlier work
Before the biotech engagement I architected and shipped several LLM-powered SaaS MVPs from zero: an AI consultation marketplace (Telegram bot + Flutter app), Contento(an LLM-driven script generator with prompt orchestration), and a Telegram-integrated e-commerce platform with payments, logistics, and warehouse integrations. Sole engineer on all of them — design, implementation, integration, release. Stack: FastAPI + PostgreSQL + Celery backends, Vue/Nuxt frontends, prompt engineering and LLM integration end-to-end.
I also built an AI-augmented multi-strategy trading system in Python + Docker — backtested across 53 crypto pairs over three years, with walk-forward adaptive strategies and per-asset trend filters.
What I'm looking for
Conversations with founders or CTOs who are putting AI into a production system and want a second pair of eyes on the architecture. Long-form consulting engagements where the work is to design what should be built, not to type lines into a file. Senior AI architecture roles at companies where engineering quality is taken seriously and "ship it" is the second sentence, not the first.
M.Sc. Computer Science, Novosibirsk State University (2001–2007) — top Russian CS program. Based in Uruguay (UTC−3) with US time-zone overlap. Available remote.