AI-powered legal document analysis platform that automatically identifies conflicts between vendor contract submissions and Massachusetts state requirements.
The system provides a React-based interface with session management, real-time progress tracking, and automated redlined document generation.
Key Demo Flow: Upload vendor contracts → Step Functions workflow orchestrates AI analysis against MA legal requirements → Download redlined documents with conflict annotations
-
🔍 Automated Conflict Detection: AI-powered analysis using Claude 4 Sonnet to identify conflicts between vendor contract language and Massachusetts legal requirements
-
📝 Intelligent Document Redlining: Automatically generates marked-up documents highlighting specific conflicts, modifications, and legal issues
Creates downloadable redlined versions with detailed conflict annotations and rationale. -
🏛️ Massachusetts Legal Compliance: Specialized knowledge base containing MA state requirements, ITS Terms & Conditions, EOTSS policies, and procurement regulations
Uses RAG (Retrieval-Augmented Generation) for accurate legal context and analysis. -
📊 Session-Based Workflow: Organize document reviews into tracked sessions with complete audit trails and historical results
Real-time progress tracking via WebSocket integration for Step Functions workflow execution (2-5 minute AI analysis tasks).
Note: The architecture diagram shows the complete system flow. The document analysis process is orchestrated by AWS Step Functions, which coordinates all Lambda functions in a multi-stage workflow.
The system follows a modern serverless microservices architecture on AWS, orchestrated by AWS Step Functions:
[React SPA] → [CloudFront CDN] → [API Gateway + WebSocket API] →
[Step Functions State Machine] → [Lambda Functions] →
[Bedrock AI + Knowledge Base] → [OpenSearch + S3 + DynamoDB]
Key Components:
- Frontend: React SPA with real-time WebSocket integration for Step Functions progress tracking
- Authentication: AWS Cognito with OAuth 2.0 flows
- Workflow Orchestration: AWS Step Functions - Orchestrates the entire document analysis workflow (11 stages)
- AI Engine: AWS Bedrock with Claude 4 Sonnet and Knowledge Base RAG (invoked by Step Functions)
- Compute: AWS Lambda functions (12 functions) orchestrated by Step Functions state machine
- Storage: Multi-tier S3 architecture with OpenSearch Serverless vector database
- Infrastructure: AWS CDK with modular construct-based deployment
| Layer | Tools & Frameworks |
|---|---|
| Frontend | React 18, React Router, Custom WebSocket Service |
| Workflow Orchestration | AWS Step Functions - Multi-stage state machine for document processing |
| Compute | AWS Lambda (Python 3.12) - 12 functions orchestrated by Step Functions |
| API Layer | API Gateway (REST), WebSocket API for real-time updates |
| AI/ML | AWS Bedrock (Claude 4 Sonnet), Knowledge Base, Titan Embeddings |
| Storage | S3 (3-tier), DynamoDB, OpenSearch Serverless |
| Auth | AWS Cognito User Pool, JWT Tokens, OAuth 2.0 |
| Infra/DevOps | AWS CDK (Python), CloudFront, IAM, CloudWatch |
| Real-time | WebSocket API for Step Functions progress tracking |
- AWS CLI configured with appropriate permissions
- Node.js 18+ for frontend development
- Python 3.12+ for CDK deployment
- AWS CDK v2 installed (
npm install -g aws-cdk)
# Clone the repository
git clone https://github.com/The-Burnes-Center/one-L.git
cd one-L
# Set up Python virtual environment
python3 -m venv .venv
source .venv/bin/activate # On Windows: .venv\Scripts\activate.bat
# Install CDK dependencies
pip install -r requirements.txt
# Configure constants (REQUIRED before deployment)
# Edit constants.py to set your stack name and Cognito domain
# Prod branch: STACK_NAME = "OneL-Prod"
# Deploy the infrastructure (creates Step Functions state machine + all Lambda functions)
cdk bootstrap # One-time setup per AWS account/region
cdk deploy
# Build and deploy frontend (handled automatically by CDK)
cd one_l/user_interface
npm install
npm run buildWhat Gets Deployed:
- Step Functions State Machine: Orchestrates the 11-stage document analysis workflow
- 12 Lambda Functions: Each handles a specific stage in the Step Functions workflow
- API Gateway: REST API and WebSocket API endpoints
- Storage: S3 buckets, DynamoDB tables, OpenSearch Serverless collection
- Knowledge Base: AWS Bedrock Knowledge Base with vector embeddings
- Frontend: React app hosted on S3 with CloudFront CDN
The system automatically generates runtime configuration post-deployment. No manual .env setup required.
| Module | Description |
|---|---|
one_l/one_l_stack.py |
Main CDK stack orchestrating all AWS resources with dependency management |
one_l/agent_api/functions/stepfunctions/stepfunctions.py |
Step Functions construct - Creates state machine and all 12 Lambda functions for workflow orchestration |
one_l/agent_api/functions/stepfunctions/*/lambda_function.py |
12 Lambda functions - Each handles a specific stage (Initialize, Split, Analyze, Retrieve, Identify, Merge, Redline, Save, Cleanup, HandleError, StartWorkflow, JobStatus) |
one_l/agent_api/agent/model.py |
Claude 4 Sonnet integration with sophisticated legal prompting |
one_l/agent_api/agent/tools.py |
Document redlining and DynamoDB operations for analysis results |
one_l/agent_api/agent/prompts/ |
AI prompts and output models for conflict detection and analysis |
one_l/agent_api/functions/knowledge_management/ |
S3 operations, Knowledge Base sync, and session management |
one_l/agent_api/functions/websocket/ |
Real-time communication handlers for Step Functions progress tracking |
one_l/user_interface/src/ |
React frontend with session management and real-time Step Functions updates |
The system employs a sophisticated multi-stage AI analysis workflow orchestrated by AWS Step Functions:
- 📄 Document Ingestion → Upload vendor submissions and reference documents with session-based organization
- 🔍 Knowledge Base Sync → Vector embedding and indexing using Titan Embed Text v2 for reference documents
- 🚀 Workflow Initiation → StartWorkflow Lambda triggers Step Functions state machine
- 📑 Document Splitting → Chunks large documents for parallel processing
- 🧠 Structure Analysis → Analyzes document structure and generates adaptive KB queries
- 🔎 Knowledge Retrieval → Retrieves relevant context from knowledge base for all queries
- ⚖️ Conflict Detection → Claude 4 Sonnet identifies contradictions, modifications, omissions, and reversals
- 🔀 Result Merging → Combines parallel chunk results into unified analysis
- 📝 Document Redlining → Generates marked-up documents with detailed conflict annotations
- 💾 Results Storage → Structured conflict data saved to DynamoDB with session tracking
- 🧹 Cleanup → Removes temporary processing files
- Zero Trust Architecture: Every service interaction requires authentication and authorization
- End-to-End Encryption: Data encrypted in transit (TLS) and at rest (S3/DynamoDB encryption)
- Fine-Grained IAM: Least privilege access with service-specific roles and policies
- User Authentication: AWS Cognito with strong password policies and JWT token management
- Audit Trail: Complete session tracking and analysis history for compliance
- No Data Retention: Documents processed securely with configurable retention policies
All implementation details are documented in the local documentation files below. For additional context or historical reference, see the external implementation guide (if available).
| Guide | Description |
|---|---|
| Deployment Guide | Step-by-step infrastructure deployment |
| Architecture Overview | Technical system architecture details |
| API Documentation | REST and WebSocket API reference |
| Command | Description |
|---|---|
cdk ls |
List all stacks in the app |
cdk synth |
Synthesize CloudFormation template |
cdk deploy |
Deploy infrastructure to AWS |
cdk diff |
Compare deployed stack with current state |
cdk destroy |
Remove all AWS resources |
npm run start |
Run React frontend locally (from user_interface/) |
Pull requests are welcome. For major changes, please open an issue first.
MIT License – see LICENSE for details.
- Built by Ritik Bompilwar, Divya Hegde, Ashley La Rotonda, and Neha Tummala
- Developed for AI4Impact initiative
- Powered by AWS serverless technologies