Principal AI Engineer · Multi-Agent Systems · Voice AI · Enterprise Automation
I build production-grade AI systems that run in real enterprise environments — regulated, integrated with legacy stacks, and held to SLOs, not just demo standards. Currently focused on multi-agent orchestration, voice AI, and agentic document processing for automotive and industrial clients.
Based in Germany. Building at wojciech UG.
Voice AI Agents — Production telephony bots with sub-second end-to-end latency. WebRTC/SIP integration, multi-turn conversation state, real-time streaming with backpressure handling and fault isolation. Handles inbound/outbound calls for automotive retail: vehicle lookup, CRM lead creation, call routing. Built on LiveKit, ElevenLabs, and OpenAI.
Agentic Document Automation — PDF intake → OCR/text layer routing → multi-step LLM processing → structured, editable output matching legally required German vehicle documentation. Deterministic rule core with LLM-powered extraction, containerized deployment, SLO-based alerting.
Multi-Agent Orchestration Platform — Unified ChannelAdapter pattern across web, telephony, SMS, and WhatsApp. Tool registries via internal npm packages. Architecture built on Mastra and LangChain/LangGraph with PostgreSQL/pgvector. Formal ADRs for every major design decision.
Languages Python · TypeScript · C#
AI/ML LangChain · LangGraph · Mastra · Semantic Kernel · LlamaIndex
Voice LiveKit · ElevenLabs · Telnyx · WebRTC/SIP
Backend FastAPI · Next.js · Node.js
Data PostgreSQL · pgvector · BigQuery · Supabase
Infrastructure Docker · Vercel · Azure · GCP
Patterns Multi-agent orchestration · RAG · Event-driven · RESTful APIs
25 years across the stack. Started coding at 12. Sitecore MVP 2008 (one of the first globally). Three years in international tech pre-sales at Sitecore across EMEA. Co-founded an AI startup (transformers/NLP). Led Data & AI at Macaw, building and managing a 6-person engineering team serving enterprise clients including Siemens, Knorr-Bremse, and Metro Digital. ML work since 2013.
I've shipped software for every layer — from CMS architectures and .NET enterprise systems to ML forecasting algorithms and real-time voice agents. The through-line is building things that work in production, not just in notebooks.
Short feedback cycles over speculative roadmaps. Readable, evolvable code over clever complexity. Constraints sharpened before scope expands. Working software over slideware.
I think like an engineer, communicate like a human, and have enough deployment scars to know where the real problems hide.
- Composable agent patterns beyond monolithic orchestration
- Claude API tool use and function calling for enterprise agent workflows
- Voice AI latency optimization at the transport layer
- Evaluation frameworks for production agent systems