AI-Tax Legal Assistant | Łukasz Augustyniak

Case studyWrocław University of Science and Technology

The challenge

Legal professionals lost hours extracting answers from millions of unstructured tax documents — court judgments, interpretations, and acts with no reliable search.

Built with

FastAPI
LangChain
SLMs (Llama, Mistral, Bielik)
LLM Pre-annotation
Human Annotation Guidelines
Elasticsearch
PostgreSQL
Claude + OpenAI

The evaluation methodology behind this work is public: LEPISZCZE, the Polish NLP benchmark published at NeurIPS 2022.