Legal Contract Risk Analyzer

A fully functional MVP application using Django that provides ML-based legal contract analysis, clause detection, deterministic extractive summarization, and heuristic risk scoring. Built for the CUAD dataset model pipeline.

Features

• PDF Extraction: Upload a PDF and extract text while retaining page numbers using PyMuPDF.
• Contract Summarization: Generates an extractive, plain-language summary based on legal keywords and clause detections.
• NLP Clause Detection: Uses a pre-trained ML model (Linear SVM via OneVsRest) to detect clauses in document chunks.
• Heuristic Risk Engine: Calculates a custom risk score and risk level based on the severity of the detected clauses.
• Evidence Referencing: Highlights exactly which page and text triggered a risk or summary section.
• Retraining Workflow: Includes a dedicated management command to safely retrain the model and save new artifacts without crashing the live server.
• Secure Storage: Stores all metadata, extraction history, and analysis runs inside NeonDB (or local PostgreSQL/SQLite fallback).

Installation

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Environment Variables

Copy the .env.example file to .env:

cp .env.example .env

Configure your variables:

SECRET_KEY=replace-this-secret-key-in-production
DEBUG=True
DATABASE_URL=postgresql://username:password@host/database?sslmode=require
ALLOWED_HOSTS=127.0.0.1,localhost

Note: If DATABASE_URL is empty, it safely falls back to a local SQLite database for development.

Database Setup

python manage.py makemigrations
python manage.py migrate

Run the Application

python manage.py runserver

Open http://127.0.0.1:8000/ to use the analyzer.

Retrain the NLP Pipeline

You can securely retrain the active model with a new dataset by using the training command:

python manage.py train_model --data data/CUAD_combined.csv --output models/

This command will:

1. Parse and melt the dataset.
2. Train Candidate Models (Linear SVM).
3. Validate artifacts.
4. Replace the live models only on success.
5. Save a training_report.json with F1 metrics.

Limitations & Disclaimers

• Text-based PDFs only: Scanned/image-only PDFs require OCR (not included).
• Not Legal Advice: This is a heuristic tool. The generated risk scores and summaries do not constitute legal advice.
• Extractive Summary: Summarization extracts sentences based on density of legal keywords; it does not "understand" legal intent.
• Storage: Uploaded contracts are stored in the server's configured storage and database. Do not upload strictly confidential information unless deployed on a compliant environment.