Agentic AI Engineering
Multi-agent orchestration, RAG pipelines, voice AI (SIP integration), LLM workflow automation. Designed, implemented, and shipped to production, not prototyped in notebooks.
- Go
- TypeScript
- Python
- Multi-Agent Systems
- Voice / SIP
Ümit Cebi — Principal AI Architect & Technical Ex-Founder
Multi-agent platforms, voice AI, and the developer tooling that keeps them maintainable. 15+ years from founder's chair to Principal seat.
Currently: Principal AI Architect at Mehrwerk GmbH Creator of Ma'at (open source) Ex-founder, Xebi Media (bootstrapped, 2014–2022 full-time)
The problem. Inbound customer calls consumed skilled staff on routine, repetitive requests, while the underlying data lived in core systems no off-the-shelf voice bot could touch.
What I built. I architected and implemented a closed multi-agent ecosystem: inbound SIP calls are routed to a voice agent, which delegates to specialized domain agents under a supervisor model. The platform integrates directly with the company's core software and triggers automated downstream workflows. Not a demo, a system handling real customers in production.
My role. End-to-end ownership: AI product strategy, architecture, hands-on implementation, and leading a 5–7 engineer team through delivery, with architecture direction across the wider 15-developer organization. Along the way I modernized the internal SDLC and introduced speed-focused delivery conventions.
4500+
calls handled per month
1350+
calls/month fully automated
88.4k€ p.a.
~335h p.m. saved
AI coding agents are users too. Ma'at is an open-source docs-as-code convention. A single Go binary plus GitHub Action, that makes AGENTS.md the canonical source of truth in a repository, generates adapters for Claude, Gemini, Cursor, Windsurf, and Copilot, and validates drift in CI.
Built because I kept watching agent instructions rot in every repo I touched.
View on GitHubMulti-agent orchestration, RAG pipelines, voice AI (SIP integration), LLM workflow automation. Designed, implemented, and shipped to production, not prototyped in notebooks.
Eight years as a bootstrapped founder taught me P&L, budgets, and C-level communication the unforgiving way. Today I use that as a Principal: translating business goals into architecture, and architecture into decisions executives can act on.
I solder flight controllers and I've poured concrete foundations. Not a skills claim — a mindset: physical systems punish sloppy dependencies instantly, and I hold software architecture to the same standard. Respect the base layer. Don't build on shifting ground.
2022 – today
Principal AI Architect / Product Owner
Company-wide AI-first roadmap, production multi-agent platform (see case study above), hands-on engineering leadership.
2014 – 2022
Founder & Managing Director
Bootstrapped media-tech agency; software and platform projects for enterprise customers, full commercial and delivery ownership. Since 2022, Xebi continues as my incubator for open-source work and product experiments. See xebi.io.
2019 – 2022
Consultant Digital Media
Platform modernization and technical feasibility for corporate clients.
2016 – 2018
Consultant & Trainer
Digital transformation consulting and training.
Occasional notes on building systems — digital and physical.
The exact same law applies to enterprise database migrations. If your atomic sync boundaries or swap patterns are flawed, the entire microservice layer above it will eventually fracture.
Read the entryTechnical product leadership means respecting the smallest integrations. A poorly documented, un-cached API endpoint breaks cluster performance just as effectively as a bridged trace on a circuit board.
Read the entry