Aosenuma Electronic Invoicing System

Building and Shipping a Complete Product

I owned this project end-to-end: from identifying the problem, designing the system, building the MVP, and deploying it to production with real invoices and real users.

The Problem I Identified

The company was manually processing incoming invoices—extracting vendor details, line items, amounts, and dates by hand. The process was slow, error-prone, and expensive. I discovered the cost inefficiency myself and pitched the solution.

The Architecture

Frontend: React interface for human review and approval of extracted data, with clear flagging of low-confidence extractions.

Backend: FastAPI service managing the extraction pipeline, validation, and audit trails.

AI Engine: Azure Document Intelligence for structured document understanding, fine-tuned to handle variable invoice formats from multiple vendors.

Data Layer: Supabase for reliable, auditable storage of extraction results and processing history.

Human-in-the-Loop Design

The system didn’t replace humans—it augmented them. Azure extracted data with confidence scores. Users reviewed high-uncertainty extractions before submission. This hybrid approach eliminated manual data entry while maintaining accuracy and compliance.

MVP Delivered and Live

I built and shipped this as a fully functional MVP that went straight into production. The system processes real invoices daily, with zero rework needed post-launch.

95% accuracy on first-pass extraction across diverse invoice formats from multiple vendors.

Measurable impact: Manual processing required 2-3 hours per day. The MVP reduced this to 30 minutes per day for human review. At $25/hour labor cost, this saves $1,500 per month in direct labor. Including error reduction and faster processing, total monthly savings: $1,800 to $2,000.

In production: Handles thousands of invoices reliably with no data loss or audit failures.

What This Taught Me

Production systems aren’t just about perfect accuracy—they’re about reliability under real constraints. I learned to build for edge cases (blurry scans, unusual formats, data gaps), design fallback flows when AI confidence dropped, and work with non-technical stakeholders who needed assurance the system worked.

The project scaled because I designed for human supervision, not full automation.