Let's Talk
I design, build, and consult on AI systems – specifically around LLM integration, data pipelines, and hybrid inference architectures. From architecture review to working implementation.
I work best with technically capable teams, in small units with short decision cycles – the kind of setup where things actually ship.
If something on this site resonates with a problem you are working on, reach out:
Selected Projects
Agentic Firmenbuch
A live data product that makes Austria's entire company register queryable by AI agents in plain language. Official master data, annual accounts, and key ratios for every firm in the country (~200k+ entities) – served over a multi-tenant MCP server, so any MCP client (Claude, Claude Code, Microsoft Copilot) answers on real, official numbers instead of hallucinating them.
Behind it sits a deterministic Azure pipeline built on the free EU High Value Datasets. It ingests every Jahresabschluss for the whole company universe, consolidates per company, and computes ratios, growth, and trends across a numbered raw → parsed → consolidated → derived → presentation flow. The raw layer is kept forever as an immutable source of truth; everything downstream is re-derivable. Version 1 is facts and clean derivations only – no scoring, no AI summaries – with the full chain verified end to end on live data.
Contract Analysis Engine
A system that takes a legal contract as a PDF and returns structured data – parties, clauses, obligations, risks. The interesting part is not what it extracts but where it runs: the same pipeline works with an 8B model on a MacBook Air and with a 54B legal-domain model on a high-end cloud GPU.
Switching between a local model and a remote one is a one-click decision in the frontend. Local inference runs through Ollama, remote inference through vLLM on a Scaleway GPU instance I provisioned and configured for this. The pipeline does not care which model is behind it – the model is a swappable component.
Everything else – extraction, normalization, chunking, structured output – lives in the pipeline around it. That is where most of the engineering effort sits, and where most of it should sit.
Business Discovery Platform
A system that aggregates businesses for sale across 6+ European marketplaces, normalizes the messy, multilingual listings into a unified schema, and enriches them using the OpenAI API for classification and summarization. The interesting part is the pipeline: a 5-layer event-driven ETL architecture where each stage triggers the next automatically via change feeds – no orchestrator, no manual intervention.
The collection runs fully autonomously – scheduled containers ingest new listings several times a day, detect changes to existing ones, and flag removals. The enriched data feeds into a semantic search layer with vector indexing and reranking, so users can query in natural language instead of filtering through forms. Three separate frontends serve different user personas.
Corporate Registry Intelligence
A system that ingests Austrian business registry filings via SOAP API, transforms the raw XML into a normalized company model, and runs every entity through a deterministic scoring engine – ownership structure, financial health, behavioral signals. No ML involved in the scoring – every rule is explicit, every score is explainable. The architecture is fully serverless, so it scales without infrastructure management.
The architecture is event-driven: each processing stage writes to Cosmos DB, which triggers the next stage automatically. The result is a dataset that surfaces acquisition and succession signals for PE firms and M&A advisors.
Before working independently, I worked across corporate (critical infrastructure), consulting, and startup environments. For more, see my About page.