TRACE
An agentic retrieval system connecting financial data, regulatory evidence and reproducible analysis.
Python · DuckDB · LangChain · Hybrid retrieval · Evaluation
The problem
Transfer-pricing analysis combines structured financial data with regulatory text. An answer needs a traceable connection to both the underlying calculation and the retrieved evidence.
The implementation
TRACE — Transfer pricing Regulatory Agentic Compliance Engine — is the research artifact of my bachelor’s thesis at TUM. It connects structured queries and text retrieval with an agent and evidence-linked reports.
Architecture
- Ingest Excel workbooks and regulatory documents.
- Store structured data in DuckDB and prepare text retrieval.
- Route queries to structured, semantic or hybrid retrieval.
- Use a ReAct agent to orchestrate analysis tools.
- Run defined compliance calculations.
- Produce structured results and evidence chains.
Research contribution
The system is compared against a rule-based baseline through a controlled ablation study. The portfolio describes the implemented architecture; quantitative evaluation claims belong alongside the exact dataset, configuration and measured results.
Repository
Explore the research repository for the implementation, setup and documentation.
The repository declares a proprietary license. Its documented test count has not been reverified as part of the emailAmigo release. This page makes no new accuracy, cost or speed claims.