Accuron AI — Invoice Parser POC

Automated Invoice PDF to Structured Excel pipeline
Handles both digital and scanned PDFs with deterministic regex-based extraction

---

What It Does

Processes invoice PDFs and extracts structured data into a multi-sheet Excel workbook:

Workflow Summary

1. Input: Digital or scanned invoice PDF files.
2. Output: Multi-sheet Excel workbook (.xlsx) containing:
   • Summary Dashboard: Overview metrics for all processed files.
   • Individual Invoice Details: Tabular detail reviews for each document.
   • details to be captured (ERP Guide): Accuron column reference rules.
   • upload format (ERP Journal Entry): Balanced double-entry 56-column journal ledger.

Fields Extracted

• Invoice Number, Date, Due Date
• Seller & Buyer details (Name, GSTIN, Address, State Code)
• Line items (Description, HSN/SAC, Qty, Rate, Amount, GST)
• Tax breakdown (CGST, SGST, IGST with rates)
• Grand Total, Round-off, Amount in Words
• Bank details, IRN, e-Invoice Acknowledgement
• Place of Supply, Reverse Charge flag

---

Pipeline Architecture

The invoice processing pipeline runs sequentially through the following modules:

1. PDF Input: Receives incoming digital or scanned document uploads.
2. Classifier: Checks for native text layers to decide the parsing route.
3. Extractor: Extracts raw text and table contents using pdfplumber (for digital text) or Tesseract OCR (for scanned pages).
4. Parser: Translates extracted text to structures via regex heuristics or Google Gemini Vision AI.
5. Validator: Runs GSTIN format checks and mathematical balancing tests.
6. Excel Generator: Writes a professionally styled, multi-sheet workbook using openpyxl.

Design Decisions

Decision	Why
Hybrid Approach (Regex + Vision AI)	Fast regex parsing for standard layouts, fallback to Gemini 1.5 Flash Vision API for complex/unstructured tables
Deterministic Base	Core extraction remains deterministic and auditable; AI only handles the heavily unstructured tables
Tesseract over PaddleOCR	30MB vs 1.5GB install, 0.5s vs 3-5s/page, production-friendly
pdfplumber over PyPDF	Superior table extraction, word-level bounding boxes
Field-level confidence	Critical fields (invoice#, date, total) carry individual scores
Error isolation	One failed PDF never blocks other invoices

---

Setup and Installation

Follow these steps to set up the invoice parser locally or deploy it to a hosting server.

1. Prerequisites (System Libraries)

Since the pipeline uses Tesseract for scanned PDF OCR, you need to install the Tesseract system binary.

• macOS (using Homebrew):

  brew install tesseract

• Ubuntu/Debian:

  sudo apt update
  sudo apt install tesseract-ocr libtesseract-dev

• Windows: Download and install the Tesseract installer from the UB Mannheim project, and add the installation folder to your Windows System PATH.

2. Local Installation

Clone the repository and install the Python dependencies:

git clone https://github.com/Punya23/AI_Invoice_Parser.git
cd AI_Invoice_Parser

# Set up a virtual environment
python -m venv venv
source venv/bin/activate  # On Windows, use: venv\Scripts\activate

# Install requirements
pip install -r requirements.txt

3. Running the Streamlit Web Interface

To launch the web dashboard locally:

streamlit run app.py

Once running, open your browser to http://localhost:8501.

• Configuring Vision AI (Optional): If you wish to use Google Gemini Multimodal parsing, enter your Gemini API Key in the sidebar input box of the web app, or set the environment variable:

  export GEMINI_API_KEY="your-api-key"

  If no key is configured, the app will automatically fall back to the offline local parser.

4. Running the CLI Orchestrator

To parse documents from the command line:

# Parse a single file
python main.py --input path/to/invoice.pdf --output output/result.xlsx

# Parse a directory of files
python main.py --input path/to/invoices_folder/ --output output/result.xlsx

# Parse files while skipping the ERP journal sheets
python main.py --input path/to/invoice.pdf --output output/result.xlsx --no-journal

5. Deployment to Streamlit Community Cloud

The codebase is fully structured to deploy directly to Streamlit Community Cloud:

1. Push your repository to your GitHub account.
2. Visit share.streamlit.io and log in with GitHub.
3. Click New app and choose the AI_Invoice_Parser repository.
4. Set the Main file path to app.py and choose Python 3.12 or the default version.
5. In Advanced Settings, add your API key under Secrets:

   GEMINI_API_KEY = "your-api-key-here"

6. Click Deploy. The server will automatically read packages.txt to install the Tesseract system binaries and requirements.txt to install Python packages.

---

Project Structure

accuron-invoice-parser/
├── main.py                    # CLI entry point
├── app.py                     # Streamlit web interface
├── requirements.txt           # Python dependencies
├── README.md
├── src/
│   ├── __init__.py
│   ├── models.py              # Data models (Invoice, LineItem, etc.)
│   ├── pdf_detector.py        # PDF type classification (digital/scanned)
│   ├── text_extractor.py      # Digital PDF extraction (pdfplumber + clustering)
│   ├── ocr_extractor.py       # Scanned PDF extraction (Tesseract OCR)
│   ├── vision_parser.py       # Multimodal Vision AI extraction (Gemini 1.5 Flash)
│   ├── invoice_parser.py      # Regex-based field extraction engine
│   ├── validator.py           # GSTIN, math, date, completeness validation
│   └── excel_generator.py     # Multi-sheet Excel workbook generator
├── Document/                  # Sample invoice PDFs (test data)
└── output/                    # Generated Excel reports

---

Technical Details

PDF Classification

Every PDF is classified before processing:

• Digital: Has embedded text layer → pdfplumber (fast, accurate)
• Scanned: Image-only → Tesseract OCR with image preprocessing
• Hybrid: Mix → digital pages via pdfplumber, scanned via OCR

Two-Tier Table Extraction

1. Tier 1: pdfplumber.extract_tables() — works on bordered tables (CIEL HR, Vault Infosec)
2. Tier 2: Positional word clustering — fallback for borderless tables (Green Clean, Tally-style invoices)
   • Groups words by Y-coordinate proximity into rows
   • Identifies column boundaries from X-position clustering
   • Reconstructs a 2D table grid

Regex Engine

• 8-10 pattern aliases per field (handles diverse invoice formats)
• dateutil.parser as fallback for date parsing
• Dynamic table column mapping based on header keyword detection
• Tally-format parser for split-line invoice numbers

Validation Layer

1. GSTIN validation: 15-character format check
2. Mathematical cross-checks: qty × rate ≈ amount, Σ items ≈ subtotal, subtotal + tax ≈ total
3. Date sanity: Parseable, not future-dated, not >5 years old
4. Completeness: Required fields present (invoice#, date, total, seller)

OCR & AI Engine (Modular)

Currently uses Tesseract for basic scanned text, and Gemini 1.5 Flash Vision API for handling complex table structures that the heuristics miss. Architecture supports drop-in replacement with:

• Google Document AI (cloud-based, highest accuracy)
• AWS Textract (production alternative)

---

Test Results

Tested against 12 Accuron invoices (5 digital + 7 scanned):

Invoice	Type	Invoice #	Date	Total	Status
AWS IT	Digital	AIN2526001124876	02/07/2025	₹27,12,845.83	Pass
CIEL HR	Digital	IHR030932627	28/04/2026	₹12,90,543.66	Pass
INUBE IT	Digital	25-26/463	11/02/2026	₹21,24,000.00	Pass
Green Clean	Digital	GC/26-27/251	17/04/2026	₹1,83,900.00	Pass
Vault Infosec	Digital	VIIPL-2627-002	02/04/2026	₹1,77,000.00	Pass
Tata Power	Scanned (OCR)	Extracted	Extracted	Extracted	Warning
OEC Records	Scanned (OCR)	Extracted	Extracted	₹1,44,356.71	Warning
Professional Couriers	Scanned (OCR)	MAA30080062	Extracted	₹974.32	Pass
Casa 2 Stays	Scanned (OCR)	BR/2526/01744	Extracted	Extracted	Warning
Saanvi Trading	Scanned (OCR)	ST/25-26/001	Extracted	Extracted	Pass

Warning = Partially extracted (OCR accuracy varies with scan quality)

---

Dependencies

Package	Version	Purpose	Size
pdfplumber	≥0.11.0	Digital PDF text & table extraction	2MB
PyMuPDF	≥1.24.0	PDF rendering for OCR (no poppler)	15MB
openpyxl	≥3.1.0	Excel workbook generation	5MB
python-dateutil	≥2.8.0	Date parsing	300KB
pytesseract	≥0.3.10	Tesseract OCR Python wrapper	50KB
Pillow	≥10.0.0	Image preprocessing for OCR	10MB
streamlit	≥1.40.0	Web interface	30MB

Total install: ~60MB (vs ~1.5GB with PaddleOCR)

System requirement: tesseract binary (brew install tesseract)

---

Production Scaling Notes

This is a POC. For production deployment:

1. OCR upgrade: Swap Tesseract for PaddleOCR or Google Document AI for higher accuracy
2. Async processing: Queue-based architecture for batch invoice processing
3. Template learning: Auto-detect invoice templates by vendor for improved regex targeting
4. API layer: FastAPI wrapper for REST endpoint integration
5. Database: PostgreSQL for invoice storage and deduplication
6. Monitoring: Field-level confidence dashboards for extraction quality tracking

---

Author

Punya Surana
Built for Accuron AI Technologies Pvt Ltd — Internship Assignment

---

Built using Python and custom Hybrid Extraction Pipelines