The problem
Refund requests arrived by email with attachments in different formats. Operators had to identify the request, extract the relevant fields, compare them with internal rules, and decide whether a human reviewer was required.
The system response
I engineered an intelligent workflow automation system that processes toll refund requests from initial email capture through final approval decision with minimal human intervention.
Key capabilities:
- Automated Gmail capture and AI classification
- OCR and NLP for document data extraction
- Historical analysis for fraud prevention
- Intelligent routing to Jira for complex cases
- Automated approval with tailored customer responses
What shipped
The workflow creates a consistent intake record, separates rule-based cases from uncertain ones, and sends review cases to Jira with the source documents attached. Customer responses remain drafts until the allowed decision path is complete.
Separates routine refund cases from uncertain ones without losing source documents or the human review point.
Architecture decisions
- Capture email and attachment references without losing the source message
- Validate extracted fields before applying refund rules
- Route uncertain or incomplete cases to a human review queue
Release checks
- Test unreadable and missing attachments
- Test conflicting extracted values and low-confidence classifications
- Confirm no customer response is released before the decision state is final
