The Invoice Intelligence Platform (IIP) is an enterprise-grade, AI-powered accounts-payable automation system. It streamlines the invoice processing lifecycle by automating data extraction, validation, anomaly detection, and ERP synchronization.
Detailed documentation is available in the docs/ directory:
- Architecture & Design - System design, tech stack, and component interactions.
- Local Setup & Deployment - Instructions for running the stack locally and deploying to Azure.
- API Reference - Documentation of the FastAPI endpoints and models.
- Contributing Guidelines - Standards for contributing to the repository.
The platform ingests vendor invoices, extracts structured data utilizing OCR and Large Language Models, and validates the extracted entities against SAP purchase orders. A machine-learning anomaly detection model evaluates the transaction to determine if human intervention is required. Approved invoices are automatically synchronized back to SAP, while analytical data is pushed to a Snowflake data warehouse for business intelligence.
- Intelligent Extraction: Combines Azure Document Intelligence with LLMs to precisely extract line items, taxes, and vendor details from unstructured PDFs.
- Automated Validation: Integrates with ERP systems (SAP) to validate PO numbers, vendor existence, and financial limits.
- Anomaly Detection: Utilizes an Isolation Forest ML model to score invoices for anomalous patterns, mitigating fraud and processing errors.
- Orchestration: Employs Azure Durable Functions for stateful, resilient workflow execution, including manual human-in-the-loop approval wait states.
- Natural Language Analytics: Features an NL-to-SQL interface powered by LLMs to query the Snowflake data warehouse using natural language.
- Frontend: React, Vite, TypeScript, Tailwind CSS
- Backend: Python 3.12, FastAPI, Pydantic v2
- Cloud (Azure): Durable Functions, Cosmos DB, Blob Storage, Key Vault, Document Intelligence, Azure OpenAI, Managed Identity
- Data & ML: Snowflake, scikit-learn, joblib
- DevOps: GitHub Actions, OpenTelemetry, Prometheus, Grafana
This project is licensed under the MIT License - see the LICENSE file for details.