Vitalii Serbyn

I've been shipping production systems for 12+ years. Started building Android apps at EPAM and GlobalLogic — one of them (Magisto) hit 100M+ users and won a Google Play award. Led teams of 2-6 engineers, owned release trains, and learned the production discipline that I now apply to AI systems.

Since 2024, I've designed and shipped 6 production AI systems through my UK company (Easelect LTD). Ascend is an orchestration daemon with 19 agents managing 4 live projects — code review, deployments, monitoring, and client reports, all trust-gated with L0-L4 policy enforcement. Crest is an AI content platform with a 6-stage LangGraph pipeline, Thompson Sampling for variant optimization, multi-model routing, and multi-platform publishing — deployed live with 91+ tests.

On the client side, I solo-architected a Web3 token launchpad on Solana (Next.js + NestJS + FastAPI, real-time blockchain indexing, 80-90% RPC cost reduction through Redis caching and bot detection) and a healthcare platform serving 10k+ users across 4 Flutter apps. PersonSearch is a 30-agent autonomous OSINT platform with dynamic sub-agent spawning. Forge UI is a 6-agent development system with 38 MCP tools for the full Flutter dev lifecycle.

Across all projects, I build the same infrastructure: multi-model routing by task complexity (GPT-4o for critical decisions, GPT-3.5 for bulk work — 70% cost reduction), full observability with Prometheus/Grafana/Jaeger/Sentry, and cost controls from day one. Every system has health checks, circuit breakers, and graceful degradation built in.

I also build my own AI development tooling. Every project runs on a custom MCP (Model Context Protocol) stack — I built semantic Code-RAG servers using ChromaDB and Sentence-Transformers for codebase-wide pattern search before writing new code. My development workflow uses multi-agent pipelines: a 9-agent system handles intake → research → spec → implementation → QA → review, with a gate system that kills 70% of feature ideas at signal validation before any code is written. Across Assisterr, 3 parallel agents coordinate across 8+ repositories using shared memory with strict namespace isolation. I don't just build AI systems — I use AI agents to build them.

I work remotely from Kyiv. Easelect LTD is structured for international contracts (UK-registered, W-8BEN-E for US clients).