Maciej Urbański
Summary
I build applied AI systems — MCP servers, agent loops, retrieval, and evals — with a confirm gate on every write. Production proof runs via a governed Wrike MCP in the system of record. Current build is a Halo ITSM MCP over a live service desk: conversation-shaped tools, not a REST wrapper, as the tool layer for a full agentic ticket loop (planner · traces · evals).
Core Competencies
- Agent systems
- MCP server & tool design · planner loops · confirm-gated writes · human-in-the-loop
- Retrieval & evals
- Agentic / hybrid / graph RAG · embeddings & chunking · Langfuse traces · tool-use evals
- Engineering
- TypeScript · Python · SQL · Zod · APIs · multi-site identity, POS & workplace systems
Tech Stack
- Languages
- TypeScript · Python · SQL · Node.js · PowerShell · Bash
- AI
- MCP (stdio) · agent loops · Zod schemas · RAG (agentic/hybrid/graph) · embeddings · Langfuse · evals
- Platforms
- Bun · Docker · Linux · Halo ITSM · Wrike · Google Workspace · AWS
Selected Projects — Agent Stack
Conversation-shaped MCP servers governing live tasks in the system of record. Narrow mutations (rename, description, status, parents), a YAML item-type registry, and a space-locked variant so changes can't drift across accounts. Ships as stdio and streamable HTTP. A person still accepts the change in Wrike.
MCP server over one live Halo instance. Tools match what a technician says — prove identity, find a ticket, load it, note it, reply to the store user — not the REST catalog. OAuth2 token cache with refresh; mutating tools preview and refuse unless confirm: true. Resolve-ticket closes the loop next.
Agent that triages tickets, tracks SLAs, and gathers the missing detail a human agent needs to resolve — plan → execute tools → stream a handoff summary. Calls Tactical RMM to run first-step troubleshooting on the endpoint before escalating, then packages findings onto the ticket. Runtime tool registry (native + MCP discovery), Langfuse-wrapped observer, SSE + session-replay JSON, abort / max-steps bounds. The agent diagnoses and enriches; a human agent resolves.
Agent framework + solved tasks: JSON-schema tool loops, MCP adapter, agentic/hybrid/graph RAG, synthetic tool-use & response-correctness evals, sandboxed code exec (QuickJS/Deno), a separate Whisper transcription MCP.
One pipeline across CLI, PySide6 GUI, and MCP. PII scan + anonymize run before any cloud LLM call; reversible AES-GCM vault; plan() is data, run() executes.
Experience
Multi-site hospitality & retail ops. Promoted Support Lead → SysAdmin → IT PM / Group Technical Lead. Own the multi-site network (identity, POS/payments, workplace) the agents act in. UAT, go-live, and hypercare decide where a tool is allowed to write.
- Google Workspace — zero-downtime identity cutover across multi-site operations; HR-driven provisioning as the standard path
- Deliverect × POS — revenue-critical vendor integration through requirements, UAT, go-live, IT/ops ownership
- Settlement reconciliation — BLIK vs POS; month-end surprises replaced by near real-time operational alerts
- Warehouse (lightweight WMS) — Next.js/React/Prisma/SQLite app live on the floor (mikr.us); catalog, search, photo cards, stock movements, CSV export; password-gated, pm2, hot SQLite backups