___  _                         ___                     _
 / __|| | __ _  _  _  ___  ___  / __|_  _  __ _ _ _  __| |
| (__ | |/ _` || || |(_-< / -_) | (_ | || |/ _` | '_|/ _` |
 \___||_|\__,_| \_,_|/__/ \___|  \___|\_,_|\__,_|_|  \__,_|

AI-Powered Contract Risk Intelligence

ClauseGuard is a production-grade agentic contract review system built on LangGraph. It autonomously analyzes any commercial contract, extracts every clause, assesses legal risk using LLM-driven judgment, and produces a structured risk report — with a human-in-the-loop checkpoint for critical findings.

---

Why ClauseGuard?

Manual contract review is slow, expensive, and inconsistent. A junior employee might miss a missing liability cap. A busy founder signs an NDA with perpetual confidentiality obligations. An auto-renewal clause goes unnoticed until the invoice arrives.

ClauseGuard addresses this by running every contract through a multi-step agentic pipeline that:

• Extracts every clause by category with no human guidance
• Applies LLM legal reasoning — not keyword matching — to assess risk
• Identifies clauses that should exist but don't
• Generates specific, actionable recommendations
• Pauses for human review when CRITICAL risk is detected

---

Agent Architecture

                          ┌─────────┐
              Contract ──►│  Ingest │ Clean text, detect contract type
                          └────┬────┘
                               │
                          ┌────▼────┐
                          │ Extract │ Extract all clauses by category
                          └────┬────┘
                               │
                          ┌────▼─────────┐
                          │ Assess Risk  │ LLM risk rating per clause
                          └────┬─────────┘
                               │
               ┌───── has_critical? ──────┐
               │ YES                      │ NO
          ┌────▼──────────────┐           │
          │ Human Checkpoint  │◄── PAUSE  │
          │  (interrupt_before)│           │
          └────┬──────────────┘           │
               │ approved=True            │
               └──────────────────────────┤
                                          │
                                 ┌────────▼────────┐
                                 │ Identify Missing │ Detect absent clauses
                                 └────────┬─────────┘
                                          │
                                    ┌─────▼──────┐
                                    │ Recommend  │ Actionable fixes
                                    └─────┬──────┘
                                          │
                                 ┌────────▼────────┐
                                 │ Generate Report │ Full RiskReport
                                 └────────┬─────────┘
                                          │
                                        [END]

---

Contract Types Supported

Type	Examples
Employment	Offer letters, consulting agreements
SaaS	Software subscription, platform licenses
NDA	Mutual and one-way non-disclosure
Vendor	Supplier and procurement agreements
B2B	Service agreements, MSAs
Other	Any unclassified commercial contract

---

Tech Stack

Layer	Technology
Agent Orchestration	LangGraph 0.2+
LLM	Anthropic Claude (claude-sonnet-4-20250514)
LLM Client	anthropic-sdk + langchain-anthropic
Tracing	LangSmith
API	FastAPI + uvicorn
PDF Parsing	PyMuPDF (fitz)
Validation	Pydantic v2
Frontend	Vanilla HTML/CSS/JS

---

Installation

git clone https://github.com/yourname/clauseguard
cd clauseguard

python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate

pip install -r requirements.txt

cp .env.example .env
# Edit .env and add your ANTHROPIC_API_KEY and LANGCHAIN_API_KEY

---

Running

uvicorn api.main:app --reload --port 8000

Then open frontend/index.html in your browser (double-click or serve with any static file server).

---

Usage

1. Open the frontend in your browser
2. Drag and drop a contract PDF or .txt file onto the upload zone
3. Watch the agent progress through each step in real-time
4. If CRITICAL risk is found, the human checkpoint panel appears — review flagged clauses, add notes, and approve or reject
5. The final risk report appears with score, clause breakdown, missing clauses, and recommendations
6. Export the report as JSON with one click

---

Sample Output

{
  "contract_type": "SaaS",
  "overall_risk_score": 87,
  "overall_risk_level": "CRITICAL",
  "executive_summary": "This SaaS agreement presents critical risk to the customer...",
  "total_clauses_found": 9,
  "clause_analyses": [
    {
      "clause": { "category": "Liability", "title": "Limitation of Liability", "text": "..." },
      "risk_level": "CRITICAL",
      "risk_reason": "No liability cap specified — Provider faces unlimited exposure.",
      "specific_concerns": [
        "Absence of financial cap on damages",
        "Customer bears asymmetric risk"
      ]
    }
  ],
  "missing_clauses": [
    { "clause_name": "Service Level Agreement", "reason": "No uptime or performance guarantees." }
  ],
  "prioritized_recommendations": [
    {
      "clause_category": "Liability",
      "issue": "No liability cap specified",
      "recommendation_text": "Negotiate a mutual liability cap of 12 months of subscription fees.",
      "priority": 1
    }
  ]
}

---

Human-in-the-Loop Checkpoint

When the assess_risk node finds any clause rated CRITICAL, the LangGraph graph pauses before human_checkpoint using LangGraph's interrupt_before mechanism. The API returns status: "awaiting_review" along with the thread ID and partial state.

The frontend renders a review panel showing all critical clauses. The reviewer can:

• Add notes
• Approve — the graph resumes and continues to identify missing clauses, generate recommendations, and produce the final report
• Reject — the graph terminates and no report is generated

Resumption is handled via POST /resume which calls graph.update_state() to inject the human decision, then graph.invoke(None, config) to continue from the checkpoint.

---

Project Structure

clauseguard/
├── agent/
│   ├── graph.py           # LangGraph state machine
│   ├── state.py           # AgentState TypedDict
│   ├── nodes/             # 6 agent nodes
│   └── prompts/           # System + user prompt builders
├── schemas/               # Pydantic models
├── ingestion/             # PDF parser + text preprocessor
├── api/                   # FastAPI app + routes
├── frontend/              # Single-file HTML/CSS/JS UI
└── tests/                 # Unit tests + sample contracts

---

Running Tests

pytest tests/ -v