// AI Systems Architect
I architect production AI agent systems
I design and ship multi-agent orchestration, LangGraph pipelines, and production AI infrastructure. 6 AI systems shipped in 18 months — from an AI content platform with a 6-stage LangGraph pipeline to a trust-gated orchestration daemon with 19 agents managing 4 live projects.
// Selected work
Systems I've shipped
Ascend: Agent Orchestration Daemon
Self-hosted orchestration daemon with trust-gated execution (L0-L4), policy engine, and audit logging. 19 agents managing 4 live projects — code review, deployments, monitoring, client reports, all automated.
- -19 agents (12 mature)
- -Trust levels L0-L4
- -Manages 4 live projects
Crest: AI Content Platform
6-stage LangGraph AI pipeline with Thompson Sampling for variant optimization, multi-model routing (GPT-4o/3.5), and multi-platform publishing. Live production deployment with 91+ tests.
- -6-stage LangGraph pipeline
- -GPT-4o/3.5 routing
- -Thompson Sampling A/B
Enterprise: Healthcare, Web3, and AI
Healthcare (Flutter/Firebase, 10k+ users), Web3 token launchpad (Solana, real-time blockchain indexing, 80-90% RPC cost reduction), and AI content platform across 9+ repositories.
- -3 production domains
- -9+ repositories
- -80-90% RPC cost cut
// What I do
Capabilities
Multi-Agent Orchestration
I design systems where AI agents collaborate autonomously — with trust layers, policy gates, and human escalation built in. Ascend runs 19 agents across 4 production projects.
- -Trust-gated execution (L0-L4) with auto-demotion on budget limits
- -LangGraph pipelines with observable state transitions
- -Custom MCP servers (Code-RAG with ChromaDB + embeddings)
- -Multi-agent dev pipelines with gate systems (70% intake kill rate)
MLOps & AI Reliability
Production AI needs quality gates and observability. I build MLflow-tracked pipelines, Thompson Sampling optimization, and full Prometheus/Grafana/Jaeger stacks.
- -MLflow experiment tracking and model registry
- -Thompson Sampling for variant optimization (150+ personas)
- -Prometheus + Grafana dashboards + Jaeger distributed tracing
- -Signal-gated feature pipelines — 5 quality gates before production
FinOps & Infrastructure
AI costs compound fast. I build multi-model routing, Redis caching layers, and per-service cost tracking. Reduced RPC costs 80-90% on a Web3 platform.
- -Multi-model routing (GPT-4o for hooks, GPT-3.5 for bodies — 70% savings)
- -Redis caching + bot detection (26K→3K daily RPC calls)
- -Docker microservice orchestration (24 services in Crest)
- -RAG pipelines with Qdrant, pgvector, hybrid retrieval
// Tools I use
Tech stack
Agent & LLM
MLOps & Platform
Backend & Data
Cloud & Infra
// About me
Vitalii Serbyn
12+ years shipping production systems — from Android apps with 100M+ users at GlobalLogic to 24-microservice AI platforms today. Solo architect across Web3, healthcare, and AI agent systems.
Now I build agent orchestration daemons, LangGraph pipelines, and production AI infrastructure with full observability. Director of Easelect LTD (UK), working remotely from Kyiv.
Book an Architecture Review
30-minute call to discuss your AI system architecture, agent orchestration challenges, or infrastructure strategy.