Grigoriy Dobryakov

Get in touch

Enterprise Architect · AI Solutions Architect · hands-on

I design systems that ship — and stay on the tools through production.

The level where the system either holds under real enterprise constraints or quietly falls apart six months in. RAG that doesn't drift, agent pipelines that don't loop, an event-driven backbone that degrades gracefully next to legacy — with a defensible reason for every trade-off, in the language of the system, not frameworks.

What I design

AI systems for production

RAG done right — chunking, re-ranking, hybrid search, knowledge graphs over corporate data. Agentic pipelines orchestrated by success criteria, not hand-written prompt chains. Hallucination and context drift diagnosed at the root: the wrong level of abstraction, not a dumb model.

Event-driven backbones

The LLM as a consumer of an event stream — Kafka, webhooks, CI/CD, monitoring. Fan-out, saga, async AI-enrichment. Safe integration next to legacy, gradual replacement without a big-bang rewrite, business continuity preserved.

Governed for production

Human-in-the-loop checkpoints, audit trails, prompt-injection defense, data isolation, GDPR / EU AI Act. Predictable behavior under real constraints — not model magic.

Artifacts

The proof. Real production systems and readable code — not sandbox demos.

Event-driven backbone at enterprise scale

Askona · $680M+ revenue, 9,000+ employees: an event-driven architecture across 40+ services and 20+ teams; RAG over the corporate knowledge base; Kafka/RabbitMQ in a real production estate.

dobryakov.net/cases/askona-digital-transformation

Public, verifiable code — NBO platform

ALS + LightGBM, Feast feature store, FastAPI/Celery, Prometheus/Jaeger observability, defined quality metrics and an A/B framework — readable end to end.

github.com/dobryakov/next-best-0ffer

My own production AI system

RAG over a knowledge base, event-driven agents, a layered knowledge architecture. The patterns I design run my own operation.

nextmoveengine.com

ML in production at scale

PersonaClick: predictive analytics and recommendation models on 199M+ customer profiles; a legacy personalization platform rebuilt for streaming / incremental learning under real performance and data-quality constraints.

dobryakov.net/cases/personaclick-benefit-segmentation

Under the hood

Stack: Anthropic Claude (API, MCP, tool use), Python, ElasticSearch, ClickHouse, SQL and vector stores, AWS / hybrid, Docker, IaC.

Built and run in production, not slideware. You can read the code, connect the MCP server, and open the running systems.

Connect me as an MCP server → dobryakov.net/mcp.html

An architect who has also led

The architecture is grounded in delivery reality because I've owned it — CTO at UMI.CMS (SDLC turnaround: revenue +50% YoY, defects 5–10/day to under 5/week, releases from months to weekly), a department turnaround with full P&L at a large system integrator, and enterprise-scale matrix leadership at Askona by architectural authority. For an architect role that's not a bid to manage — it's why the design accounts for org reality: staffing, politics, and what a team can actually run.

Let's talk.

Open to Enterprise Architect / AI Solutions Architect / System Architect roles — individual-contributor, hands-on.

Fully remote · EU-hours (CET/CEST) · open to relocating within the EU and working remotely from there (Blue Card–eligible) · EOR or contractor.

Still thinking it over? Here's a pile of links about me — look around first, write when you're ready.