01Challenge

A design company's knowledge base is hundreds of documents: standards, regulations, tables and diagrams. Requirements duplicate, refine and contradict each other, and when a new version of a document comes out there is no quick way to see which solutions and checks are affected.

02Solution

• Parsing of DOCX, PDF and XLSX into fragments with exact addresses: document, version, clause or cell
• AI extracts verifiable requirements: object, metric, value, conditions, exceptions and a verbatim quote
• Links between requirements — refines, depends on, duplicates, contradicts — and to standard solutions and checks
• A requirement or link enters the register only after an expert's decision; every change creates a new revision
• Document version comparison and impact analysis: affected solutions, checks, requirements and projects with an effort estimate
• UI: a register with export, a document view with highlighted sources, a link graph and a review queue

03Results

• On new documents: extraction precision of 94.8% and recall of 100% against targets of 90% and 85%
• On tuning documents and after refinement — 100% extraction precision and recall
• Contradictions between requirements are detected automatically and routed to an expert
• The method, a test report and a pilot-stage protocol template were delivered