← Back to project catalogue
GP-DA-03LVEUVData and AIOpen for request

MSME invoice document processing

A document-intelligence workflow for extracting, validating, and reviewing structured fields from varied Indian MSME invoice layouts.

  • Nuxt
  • Vue
  • FastAPI
  • PostgreSQL
  • Python
  • PyTorch
  • OpenCV
  • Docker

Project definition

Problem statement

MSME invoices can differ in layout, print quality, language, table structure, tax presentation, and item descriptions. Basic OCR returns text but does not reliably identify fields or explain which values require review.

The engineering problem is to combine image processing, OCR, layout understanding, field validation, and confidence scoring so that invoice data can be extracted into a consistent format with a clear human-correction path.

Project objectives

  • Process PDF and image invoices with varied layouts and scan quality.
  • Extract selected supplier, recipient, invoice, tax, item, and total fields.
  • Validate formats, arithmetic relationships, and selected GST invoice rules without making compliance decisions.
  • Use field-level confidence to prioritise uncertain values for human review.
  • Measure OCR, extraction, table, validation, and correction performance.

System design

System modules

01

Document intake

Accepts permitted PDF and image files, checks type and size, creates page images, and records processing state.

02

Image and OCR pipeline

Corrects orientation and image quality, runs OCR, and retains words, positions, confidence, and page references.

03

Layout and field extractor

Identifies key-value regions, tables, line items, and selected invoice fields through replaceable extraction methods.

04

Validation engine

Checks field formats, GSTIN structure, dates, tax calculations, totals, and required relationships under a documented rule set.

05

Human review workspace

Highlights low-confidence or invalid fields on the source page and records corrections with an audit history.

06

Export and evaluation module

Exports accepted records to a defined JSON or CSV structure and compares predictions with prepared ground truth.

Methodology

System workflow

  1. 01
    Upload invoice

    A permitted invoice file is validated, converted into page images, and assigned a processing record.

  2. 02
    Read document

    Image preparation and OCR produce positioned text with confidence values.

  3. 03
    Extract fields

    The system identifies header fields, party details, line items, tax values, and totals in the target structure.

  4. 04
    Validate and review

    Format and arithmetic checks identify uncertain or inconsistent values for correction against the source image.

  5. 05
    Accept and export

    The reviewed record is stored with its audit history and exported in the documented structured format.

Demonstration scenario

A user uploads a prepared MSME invoice containing supplier details, GST fields, line items, and totals. The system extracts the record, highlights one low-confidence field and one arithmetic mismatch, and opens the source region for correction. The accepted invoice is exported in the documented JSON structure with its review history.

Engineering

Technical architecture

Web application
Nuxt and Vue for upload, page viewing, field correction, validation messages, status tracking, and export.
Document API
FastAPI for secure file intake, processing jobs, extracted fields, corrections, validation, and export endpoints.
Data layer
PostgreSQL for document metadata, field values, confidence, validation results, corrections, and experiment records.
Vision pipeline
Python, OpenCV, OCR, and replaceable layout-analysis models for page preparation, text recognition, tables, and fields.
Validation model
A versioned set of format, consistency, arithmetic, and selected schema rules with field-level reasons.
Delivery
Docker-based setup with safe demonstration invoices, target schemas, trained or configured extractors, and reproducible tests.

Testing

Evaluation

Evaluation measures

  • Character and word error rates on prepared invoice pages
  • Field-level exact match, precision, recall, and F1 score
  • Line-item and table extraction accuracy
  • Validation-rule precision on prepared correct and incorrect records
  • Correction time and review rate at selected confidence thresholds
  • Processing time and failure rate across layouts and image conditions

System boundaries

  • The included dataset uses synthetic or properly redacted invoices and contains no live buyer or supplier records.
  • The prototype validates a documented subset of invoice fields and rules and does not submit data to GST systems.
  • A successful check is not a legal, tax, accounting, or compliance approval.
  • Human review remains required before extracted data is used in a business process.

Included

  1. 01Invoice upload, review, correction, and export interface
  2. 02OCR, layout analysis, field extraction, and validation pipeline
  3. 03Confidence scoring and human-review workflow
  4. 04Prepared synthetic and redacted invoices, experiments, and results
  5. 05Complete source code in a private GitHub repository
  6. 06Complete project documentation with synopsis, abstract, methodology, processing diagrams, extraction results, screenshots, and conclusion
  7. 07Setup and usage guide

Project record

No buyer information is collected on this page.

Permanent project ID
GP-DA-03LVEUV
Catalogued
21 Aug 2026
Completed
Pending
Verified
Pending
Demonstration
Added when ready

Handover

After purchase

  1. 01
    Payment is confirmed

    The project is marked unavailable and cannot be purchased again.

  2. 02
    Repository access is granted

    The buyer's submitted GitHub account receives access to the private repository.

  3. 03
    The purchase record is delivered

    The certification sheet is prepared from the reviewed buyer details and sent privately by email.