Intelligent Document Parser

An end-to-end backend system for extracting structured information from PDFs and images using OCR and transformer-based NLP models.

The project focuses on production-style system design rather than model experimentation. It demonstrates how machine learning components can be integrated into a scalable backend pipeline for document processing.

The system processes uploaded documents asynchronously, extracts raw text using OCR, identifies entities using a pretrained NER model, and returns structured data via a REST API.

Key Features

• Document ingestion via REST API
• PDF and image processing
• OCR-based text extraction
• Transformer-based Named Entity Recognition (NER)
• Asynchronous job processing
• Status tracking for processing jobs
• Structured JSON output
• Production-style backend architecture

System Architecture

The project is designed using a layered architecture commonly used in production ML systems.

Client
  |
  v
FastAPI API Layer
  |
  v
Service Layer
  |
  v
Async Job Queue
  |
  v
Document Processing Pipeline
    ├── Text Extraction (OCR)
    ├── NLP Processing (NER)
    └── Entity Post-processing
  |
  v
Persistence Layer

Processing Flow

1. Client uploads a document (PDF or image).
2. API stores the document and creates a processing job.
3. The job is queued for asynchronous processing.
4. Worker retrieves the job and executes the pipeline:
   • Extract text using OCR
   • Run transformer-based NER to identify entities
   • Post-process entities
5. Extracted data is stored and returned as structured JSON.

Project Structure

app/
 ├── api/
 │   └── endpoints
 ├── core/
 │   ├── config
 │   ├── logging
 │   └── security
 ├── pipeline/
 │   ├── text_extractor
 │   ├── ocr_processor
 │   ├── ner_processor
 │   └── pipeline.py
 ├── services/
 │   ├── document_service
 │   └── job_service
 ├── workers/
 │   ├── worker
 │   └── tasks
 ├── persistence/
 │   ├── database
 │   ├── models
 │   └── repositories
 ├── messaging/
 │   └── queue
 └── schemas

The architecture separates concerns between:

• API handling
• business logic
• ML pipeline
• async workers
• persistence

Technologies Used

Backend

• FastAPI
• Python
• Async workers
• REST APIs

Machine Learning

• Transformer-based NER
• OCR for document text extraction

Infrastructure

• Docker
• Redis (message broker)
• PostgreSQL (persistence)

Example Output

{
  "document_id": "123",
  "status": "COMPLETED",
  "entities": {
    "invoice_number": "INV-1024",
    "vendor": "ABC Pvt Ltd",
    "total_amount": "₹12,450",
    "date": "2024-01-10"
  }
}

Running the Project

Clone the repository

git clone https://github.com/yourusername/intelligent-document-parser.git
cd intelligent-document-parser

Install dependencies

pip install -r requirements

Run the API

uvicorn app.main:app --reload

Design Goals

This project was built to demonstrate:

• Integration of ML models into backend systems
• Asynchronous document processing pipelines
• Clean separation between API, services, and ML components
• Scalable architecture for real-world document processing workflows

Limitations

Invoice and receipt formats vary widely across organizations.
The current system focuses on demonstrating pipeline architecture and ML integration, rather than solving layout variability across all document types.

Improving extraction accuracy would require:

• layout-aware models
• specialized invoice datasets
• domain-specific model fine-tuning

Future Improvements

• Layout-aware document understanding models
• Table extraction
• Improved entity post-processing
• Multi-document type classification
• Production deployment on cloud infrastructure