I document my work, projects and research interests around applied AI, knowledge graphs, fullstack development, networking and enterprise systems.
“Geht ned? Gibts ned!” — Austrian for: there’s always a solution.
Multi-agent Knowledge Graph traversal as a grounding mechanism for LLM-based question answering — completed with the highest grade (Excellent, 1 on the Austrian scale). Read more ↓
Topics I actively work on, study or find worth understanding more deeply.
Technical and academic artefacts — each a response to a concrete question or problem.
Prototype for an AI agent-based automated invoice processing pipeline with direct SAP integration. Demonstrated at the 20th INFORMATICS SAPshots event alongside manufacturing and healthcare use cases.
Read more →
Developed with three fellow students — a two-stage AI system that identifies interviews within the APA news database and optimises search relevance. Trained exclusively on secure, private infrastructure.
View on LinkedIn →
Built a fully paperless office using paperless-ngx, with automated invoice pre-accounting via local LLMs running on Ollama. Document classification and data extraction handled through n8n workflows.
An architecture concept for a document assistant used by a few thousand people on long, structured documents — contracts, legislation, briefs. Two problems drive the whole design: keeping cost and latency flat whether a workspace holds five documents or five hundred, and sharing a document set between people and between tools without inventing a second permission model for each.
A Model Context Protocol server for Austrian Railways. It gives Claude and other MCP-capable assistants three tools — station search, journey planning and station boards — so a travel question can be answered against live timetable data instead of a training snapshot. A purely private project — not affiliated with, endorsed by or connected to ÖBB.
A home that understands situations instead of waiting for commands, and takes the small everyday decisions off your hands. The goal is not more smart home — it is less for you to think about: no setting things in the morning, no checking whether everything is switched off, no app for every part of the house, no rules for you to maintain.
Coming soon · intelligentliving.fuxiq.at
A route can work on the map and still be impassable in reality. A single broken lift is not a minor inconvenience — if it is the only step-free access to a station, the route does not get longer, it disappears. EveryWay plans journeys that already know the current state of the infrastructure and reroute around it before somebody is standing in front of the problem.
Visit the site →Multi-agent knowledge graph traversal against hallucination — at WU Vienna.
Large language models produce fluent answers that nothing supports — hallucination. This thesis asks whether an open-weight model that walks a knowledge graph with several parallel agents answers more faithfully to its evidence than three text-based approaches: one-shot prompting, two-step retrieval-augmented generation (RAG) and agentic RAG. The system splits the question into clauses, finds an entry entity for each in a vector index over the graph, lets parallel sub-agents walk outward from those entries, consulting the language model only every few hops, and answers from the best sub-agent’s subgraph. Its premise is that of the Wikipedia game: the entry need not be exact, because the walk will pick up the trail.
The evaluation has two phases on Wikidata: a medical subgraph against an external benchmark, then a skeleton of the complete graph against question sets it answers by construction, once with an entry index that contains the answer entities and once with the system’s own.
The intended outcome — more faithful at comparable accuracy and lower cost — is not reached. The first phase separates the graph system from nothing. In the second, a single-agent walk given a correct entry is more accurate than one-shot prompting and agentic RAG and about as faithful as the text baselines; with the system’s own index neither graph arm beats the ungrounded baseline, because most walks return no fact. The graph does not mislead; it is absent. The walk picks up the trail only when it starts beside it, and the multi-agent layer adds no significant accuracy at many times the cost.
The contribution is a located bottleneck — entity linking against a whole-graph index — and its repair: looking the question’s names up in the graph’s own label table, which lifts the single walk close to the diagnostic ceiling at fewer model calls. The result vindicates the traversal, not the system built around it.
Technologies and domains I work with across enterprise, AI and infrastructure contexts.
For academic exchange, technical discussions or project-related questions, feel free to reach out via LinkedIn, GitHub or Bluesky. I am also reachable by email.