Built a document understanding system for risk classification:

• OCR with layout preservation, table extraction, and heuristic cleaning.
• Hybrid CV/NLP features with lightweight model to meet inference budgets.
• Emphasis on transparency and traceability for auditors.