🕵️ AuditGraph: Autonomous Corporate Due Diligence Agent

Author: Venkata Narendra Nuka

AuditGraph is an advanced Agentic GraphRAG (Retrieval-Augmented Generation) system designed to automate forensic accounting and corporate due diligence.

It combines unstructured financial data (PDF contracts, 10-K reports, internal emails) with a Neo4j Knowledge Graph to detect complex fraud patterns such as:

• Related-party transactions
• Shell company fund transfers
• Hidden ownership structures

---

🚀 Key Features

🔗 Multi-Modal GraphRAG Architecture

• Goes beyond standard vector search by modeling explicit relationships

• Tracks:

  • Ownership structures
  • Board memberships
  • Financial transfers

• Enables multi-hop reasoning across disconnected datasets

---

🧠 Cognitive Data Engineering

• Powered by NVIDIA Nemotron-3-Super-120B for enterprise-grade reasoning
• Uses explicit Reasoning Budgets ("Thinking") to reduce hallucinations
• Processes complex PDFs using pypdf
• Converts raw text into structured JSON ontologies based on custom schemas

---

🤖 Autonomous LangGraph Agent

• Features an autonomous forensic accounting chatbot
• Dynamically generates and executes Cypher queries
• Performs step-by-step reasoning over graph evidence
• Produces explainable forensic investigation reports

---

🛡️ Robust Extraction Pipeline

• Handles inconsistent LLM outputs safely

• Uses:

  • AST parsing
  • Strict schema prompts
  • Validation mechanisms

• Ensures zero pipeline breakage during ingestion

---

📊 Interactive Dashboard

Built with Streamlit, featuring:

• Real-time Neo4j graph metrics
• Autonomous chatbot for graph querying
• Analyst-friendly interface
• Interactive investigation workflows

---

🛠️ Technology Stack

Category	Tools Used
Language	Python 3.10+
Graph Database	Neo4j AuraDB (Cloud)
LLM Engine	NVIDIA Nemotron (langchain-nvidia-ai-endpoints)
Orchestration	LangChain & LangGraph
Data Processing	PyPDF
Frontend	Streamlit

---

📂 Project Structure

AuditGraph/
│
├── app/
│   └── ui.py                  # Streamlit Dashboard UI
│
├── data/
│   └── raw/
│       ├── contracts/
│       ├── financials/
│       └── internal/
│
├── src/
│   ├── graph/                 # Neo4j connection & schema logic
│   │   ├── builder.py
│   │   ├── neo4j_client.py
│   │   └── schema.py
│   │
│   ├── ingestion/             # PDF reading & NVIDIA JSON extraction
│   │   ├── loader.py
│   │   └── parser.py
│   │
│   ├── orchestrator/          # Autonomous Agent Logic
│   │   ├── langgraph_bot.py
│   │   └── tools.py
│   │
│   └── config.py              # Environment variables
│
├── tests/
├── .env
├── main.py                    # Pipeline execution script
└── requirements.txt

---

⚙️ Installation & Setup

1️⃣ Clone Repository

git clone https://github.com/yourusername/AuditGraph.git
cd AuditGraph

---

2️⃣ Create Virtual Environment

python -m venv .venv

# Windows
.venv\Scripts\activate

# Linux / macOS
source .venv/bin/activate

---

3️⃣ Install Dependencies

pip install -r requirements.txt

---

4️⃣ Setup Neo4j AuraDB (Cloud)

1. Open the Neo4j Aura Console.
2. Create a Free instance.
3. Save the generated password.
4. Copy your connection URI.

Example:

neo4j+s://<instance-id>.databases.neo4j.io

---

5️⃣ Configure Environment Variables

Create a .env file in the project root:

NVIDIA_API_KEY=your_nvidia_api_key_here

NEO4J_URI=neo4j+s://<your-id>.databases.neo4j.io
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=your_auradb_password

---

🏃 Running the System

▶️ Step 1: Ingest Documents

Add PDF files to:

data/raw/

Run:

python main.py

The system will:

• Parse PDFs using PyPDF
• Extract structured entities via NVIDIA Nemotron
• Build the Neo4j knowledge graph
• Store relationships in AuraDB

---

📊 Step 2: Launch Dashboard

streamlit run app/ui.py

Open:

http://localhost:8501

to interact with the autonomous forensic investigation agent.

---

🧠 Use Cases

AuditGraph supports:

• Corporate due diligence
• Fraud detection
• Financial investigations
• Risk analysis
• Compliance auditing
• Related-party transaction discovery
• Ownership structure analysis

---

⚠️ Notes

• Ensure your Neo4j AuraDB instance is Running before ingestion.
• Keep your .env file private and excluded from version control.
• Large documents (100+ page 10-K reports) may require several minutes for deep reasoning and extraction.

---

🤝 Contributing

Pull requests are welcome.

For major changes, please open an issue first to discuss proposed modifications.

---

📜 License

MIT License

---

⭐ If you found this project useful, consider starring the repository!