Turning physical cash vouchers into validated accounting records.
Accounts AI connects scanning, OCR extraction, human validation, approval, history tracking, and Oracle storage in one controlled workflow.
- Role
- Primary solo ownership
- Environment
- Internal accounting workflow
- Stack
- Python · Flask · OCR · Oracle
Time reductions are workflow-based estimates from development testing. Formal organisation-wide savings have not been measured publicly.
Manual entry made every voucher a small, repetitive data task.
Staff had to read each physical cash voucher, identify multiple fields, enter the information, verify it, and preserve an accurate record history.
The challenge was not simply reading text. Documents vary in clarity and layout, so uncertain extraction needed to remain visible and correctable before any accounting record was approved.
Automation accelerates entry. People retain control.
The workflow separates machine extraction from accountable approval.
- 01
Scan
A physical voucher enters through the scanner workflow.
- 02
Extract
OCR identifies text and maps accounting fields.
- 03
Review
A person checks, corrects, and completes the record.
- 04
Approve
Only reviewed information can move forward.
- 05
Store
The approved record and history are stored in Oracle.
One review surface for the document and its structured record.
₹ 620.00
Purpose: Sample operational expense
Amount in words: Six hundred twenty only
The interface shown here uses fictional values. No client documents, identities, or production data are displayed.
Reliability comes from the system around OCR.
Structured extraction
Detected text is mapped into specific accounting fields instead of accepted as an unstructured block.
Human validation
Users can review and correct uncertain fields before approval.
Approval boundary
Extraction alone does not authorize a final accounting record.
History and traceability
Processed documents and status changes remain accessible for later review.
Primary independent engineering across the complete workflow.
I built the platform primarily independently, while domain users helped explain requirements and validate business correctness.
What this evidence does—and does not—claim.
- Performance varies with scan clarity, handwriting, and document layout.
- Human review remains necessary for uncertain or consequential fields.
- The 50–70% figure is an estimate from development workflow testing, not an audited organisation-wide result.
- Public visuals are sanitized; client identity and production information remain confidential.
Have a document workflow worth automating?
I’m available for selected document-intelligence and business-automation projects.