Built an end-to-end invoice and document processing platform for a leading accounting services provider, automating the capture, parsing, validation, and circulation of multilingual invoices and supporting documents. The system predates the generative-AI era and was engineered entirely on classical image processing, computer vision, and tightly tuned OCR — proof that disciplined pipeline engineering delivers production-grade extraction accuracy long before foundation models existed.
The Challenge
Accounting teams handle thousands of invoices, receipts, and supporting documents every month — scanned at varying quality, written in multiple languages, and laid out in every format vendors care to invent. Before LLM-based document AI made headlines, automating this work meant solving every part of it the hard way: cleaning up noisy scans, locating tables and fields without a transformer hint, reading text in scripts that off-the-shelf OCR engines weren't trained for, and turning raw recognition output into validated, posting-ready records. The partner's accountants needed all of it to work reliably enough to remove paper from the loop, in production, every day.
Our Approach
We built a complete capture-to-archive pipeline grounded in classical computer vision and tightly tuned OCR — no large language models, no foundation-model crutch. Image preprocessing handled the messy reality of scanned input: deskewing, adaptive binarization, denoising, perspective correction, and contour-based page segmentation. Layout analysis broke each document into logical zones (header, vendor block, line items, totals) using projection profiling, connected-component analysis, and template matching tuned to the invoice formats accountants see most often. The OCR layer was language-tuned to read multilingual content reliably, and a rules-and-statistics extraction stage turned recognized text into structured fields with cross-validation — line-item sums against totals, tax math, date and ID formats — so downstream systems could trust the output. The whole pipeline plugs into a workflow engine that handles intake, review, approval, and archival, replacing the manual paper trail with a searchable digital one.
Results
The platform automates the heavy, repetitive end of accounting document work, reading multilingual invoices off scans and pushing structured, validated data straight into downstream finance workflows. More broadly, the project is a record of what disciplined engineering can do without generative AI: production-grade extraction accuracy and accountant-grade reliability, built from classical image processing, computer vision, and OCR that had to be right the first time. Many of the same engineers and patterns now power our modern document-AI work — the foundations didn't expire, they compounded.
