Selected work
Work
12 Projects
iQuantum
Open-source AI coding agent for the terminal — a Plan → Implement → Validate loop routes work between a reasoning "Architect" and a fast "Editor", runs every change in an isolated Docker sandbox, and commits only once the project's tests pass.
DataHydra
Lead enrichment & verification API aggregator — cost-optimized waterfall routing tries cheap data sources first and escalates only when needed, with a unified schema, multi-source quality scoring and GDPR-aware consent tracking.
KanzleiAgent
DATEV-compatible AI co-pilot for German tax firms — Mandant document ingestion, SKR03/04 booking proposals and a Kanzlei-private RAG corpus with a hash-chained audit trail. The Steuerberater always signs off.
RechnungsRadar
E-invoice compliance & accounts-payable agent for the German Mittelstand — XRechnung/ZUGFeRD validation against EN 16931, AI classification for ERP posting and GoBD-compliant WORM archiving.
CoRAG
Multi-hop question answering through adaptive, iterative retrieval — it decomposes a complex query, retrieves, identifies knowledge gaps in the results, and refines the search until it can answer, instead of retrieving once.
Syntax-Aware-RAG
This project develops a sophisticated, syntax-aware chunking strategy that respects the semantic and structural boundaries of a document. It will use NLP libraries to chunk documents into sentences or logical sections (e.g., paragraphs, list items).
Rosetta-Transformer
Historical archives contain vast amounts of knowledge locked away in ancient or low-resource languages, often on damaged manuscripts. This project applies advanced NLP to the digital humanities by building a system to aid in the deciphering and analysis of these texts.
deeprag
This project aims to reimagine the RAG pipeline by replacing the discrete, non-differentiable retrieval step with a unified, end-to-end trainable neural network.
VideoRAG
This project aims to build a system capable of answering natural language queries over a large collection of videos. This requires a multimodal approach to indexing, where video content is processed and made searchable through a combination of visual and auditory information.
Nothing here — try another filter.