Agentic Contract Risk Analyzer

A multi-agent AI system that evaluates contracts against regulatory obligations (EU AI Act, Australian frameworks) using clause extraction, obligation mapping, RAG-based regulatory grounding, verification, and executive scoring.

This project demonstrates how Agentic AI can move beyond prompt-based analysis into structured, auditable, multi-stage reasoning pipelines.

---

🚀 What Makes This Different?

Unlike traditional GenAI contract analyzers that rely on a single LLM pass, this system:

• Uses a 5-Agent Orchestration Pipeline
• Grounds findings using regulatory PDFs via RAG (TF-IDF index)
• Filters obligations per clause type (reduces noisy cross-product analysis)
• Verifies evidence and citations before marking findings as validated
• Produces an Executive Risk Scorecard

---

🧠 Architecture Overview (5-Agent System)

1️⃣ Extractor Agent

• Splits contract into structured clauses
• Assigns clause types (SECURITY, PRIVACY_DATA, TRANSPARENCY, etc.)
• Captures evidence spans

2️⃣ Obligation Mapper Agent

• Loads curated regulatory obligations (EU AI Act, AU)
• Each obligation declares clause_types_applicable
• Reduces irrelevant comparisons

3️⃣ Auditor Agent (RAG-powered)

• Cross-checks clause vs relevant obligations
• Retrieves regulatory evidence from indexed PDFs
• Produces grounded findings with citations

4️⃣ Verifier Agent

• Enforces:
  • Contract evidence present
  • Regulatory citations present
  • Confidence threshold
• Flags findings as NEEDS_REVIEW when grounding is insufficient

5️⃣ Reviewer Agent

• Aggregates risk metrics
• Generates executive-ready Risk Scorecard
• Produces summary + next steps

---

📂 Project Structure

src/ agents/ extractor.py obligation_mapper.py auditor.py verifier.py reviewer.py rag/ indexer.py retriever.py store.py regulation/ obligations_catalog.json orchestrator.py model_adapter.py

streamlit_ui/ dashboard.py

data/ regulations/ eu_ai_act/ australia/ rag_index/

tests/

---

🏗 How to Run

1️⃣ Install dependencies

pip install -r requirements.txt

2️⃣ Add Regulatory PDFs

Place official documents in:

data/regulations/eu_ai_act/ data/regulations/australia/

3️⃣ Build RAG Index

python scripts/build_reg_index.py

4️⃣ Launch UI

streamlit run streamlit_ui/dashboard.py

Enable:

• Agentic mode
• Require regulation citations

---

🧪 Testing

Run:

python -m pytest -q

Includes:

• Applicability filtering test
• Scoring monotonicity test
• Evaluation harness
• Model adapter smoke tests

---

📊 Example Output

• Overall Risk Score
• Risk Level
• Findings with:
  • Severity
  • Confidence
  • Contract Evidence
  • Regulatory Citations (page + excerpt)
• Executive Risk Scorecard
• Needs Human Review indicator

---

🎯 Why This Matters

This project demonstrates:

• Agentic AI system design
• Deterministic + explainable architecture
• Regulatory grounding (EU AI Act)
• Multi-stage reasoning instead of single-pass LLM output
• Enterprise-ready audit trail design

---

⚖️ Disclaimer

This tool is for educational and demonstration purposes. It does not constitute legal advice. Human legal review is always required.

---

👤 Author

Madhur Khandelwal
AI Transformation | Agentic AI Systems | Regulatory-Aware AI Design

GitHub: https://github.com/MadhurKh