Verifiable LLM Agents for Trade Document Processing and Customs Compliance
Verifiable LLM Agents for Trade Document Processing and Customs Compliance
Description
Details
Context and Problem Statement
Export transactions involve heterogeneous documents such as invoices, packing lists, certificates of origin, and customs declarations. Multimodal models can extract information effectively, but fabricated values are unacceptable in regulated workflows.
Research Question
How can a multimodal agent be constrained so that no extracted value is produced without verifiable localization in the source document, and what automation rate remains achievable under this constraint?
Proposed Approach
Combine multimodal extraction with mandatory spatial grounding, a symbolic rule engine for consistency and compliance checks, and human review for unsupported assertions. Tariff classification is treated as a traceable recommendation task.
Expected Contribution
A neuro-symbolic architecture with experimentally evaluated non-fabrication behavior and a realistic measure of automation under traceability constraints.
Expected Prototype
An export-document processing application producing consistency reports, detected discrepancies, and an audit trail linking every field to its source region.
Datasets
DocVQA, FUNSD, and anonymized commercial documents from a logistics partner.
Challenges
Variable scan quality, French-Arabic-English multilingualism, evolving customs rules, confidentiality, and legal responsibility.
Research Question
Innovation
Expected Deliverable
Technologies
Required Skills
- Document Vision and Understanding
- LLM Agent Engineering
- Logic Programming or Rule Engines
- International Trade and Customs Fundamentals
Datasets
- DocVQA
- FUNSD
- Anonymized commercial documents from a logistics partner