AI-Powered Purchase Order Processing System

A global medical device manufacturer was processing purchase orders by hand — customer service staff reading emails, opening Word, Excel, and PDF attachments, and keying orders into JD Edwards. The build automated email categorization, OCR extraction, and JSON integration into the ERP. The harder problem was what the team was measuring: early on, success was framed as the share of POs processed without human touch. But a wrong quantity that clears automatically is worse than one flagged for review, so we shifted to optimize for extraction accuracy instead and to treat confidence-based manual review as a feature rather than a failure.

What I did

• Managed delivery and served as the liaison between the client's operations team and engineering through requirements, build, and rollout
• Ran acceptance testing across the Azure Copilot AI + OCR pipeline and the SharePoint validation interface
• Tuned extraction accuracy iteratively against real order documents, working with the client's team rather than directing it
• Drove estimated $210,500 in annual labor cost savings