FreeLegal AI - Freelancer Legal Assistant

FreeLegal AI
FreeLegal AI
FreeLegal AI
Architecture

Inspiration

As freelancers ourselves, we've experienced the frustration of:

Clients refusing to pay for completed work
Signing contracts with hidden unfair clauses
Not understanding legal terms like "indemnification" or "force majeure"
Losing money because we didn't know how to negotiate better terms

Hiring a lawyer costs $300-500/hour - money most freelancers don't have for routine issues. We built FreeLegal AI to democratize legal help for the 1.57 billion freelancers worldwide who deserve fair treatment but can't afford traditional legal services.

What it does

FreeLegal AI provides instant legal assistance through 4 specialized AI agents, each powered by Claude 3.5 Sonnet v2 via AWS Bedrock AgentCore:

⚡ Action Agent

Purpose: Immediate help for urgent legal issues
Example: "My client won't pay $5000" → Receives step-by-step action plan with country-specific legal options, demand letter templates, and timeline

🔍 Analysis Agent

Purpose: Contract review and red flag detection
Example: Upload a PDF contract → Identifies unfair payment terms (90-day payment vs industry standard Net-30), missing clauses, liability risks, and provides recommended changes

💡 Explanation Agent

Purpose: Translates legal jargon into plain English
Example: "What does indemnification mean?" → Receives simple explanation with real-world examples, risks, and how to protect yourself

🤝 Negotiation Agent

Purpose: Tactical advice for contract negotiations
Example: "How to push back on IP transfer?" → Receives negotiation scripts, counter-proposals, and professional talking points

Key Features:

📄 Contract Upload: PDF/TXT upload with automatic text extraction using AWS Textract
💬 Context-Aware Conversations: Agents maintain session memory across messages
🌍 Localized Advice: Country-specific legal guidance (e.g., Egyptian Economic Courts, US small claims)
🎯 Role-Based Intelligence: Each agent specializes in one domain for expert responses
⚡ 10-Second Responses: Fast enough for urgent situations

How we built it

Architecture

User Browser → CloudFront (CDN) → S3 (Static Website) ↓ API Gateway → Lambda (Orchestrator) → Bedrock AgentCore → Claude 3.5 Sonnet v2

Frontend (HTML/CSS/JavaScript)

4 specialized chat interfaces (one per agent)
Responsive design with real-time upload feedback

Backend (AWS Lambda - Python 3.12)

orchestrator_lambda_proxy.py: Routes requests to appropriate agent
Prepends role-specific instructions to differentiate agent behavior:
- Action: "Provide immediate action plans with timeline..."
- Analysis: "Review contracts for red flags, compare to industry standards..."
- Explanation: "Explain legal terms in 8th-grade English with examples..."
- Negotiation: "Provide tactical negotiation advice with scripts..."

AI Layer (AWS Bedrock AgentCore)

Model: Claude 3.5 Sonnet v2 (us.anthropic.claude-3-5-sonnet-20241022-v2:0)
AgentCore Runtime: freelancer_action_agent-Q83Rk73nkD
Memory: Session-based conversation history (freelancer_action_agent_mem-Fh4JVHDpfJ)
Deployment: Used agentcore launch CLI for agent deployment

Infrastructure

S3 Bucket: Static website hosting (freelancer-legal-assistant-frontend-897722703585)
CloudFront: Global CDN distribution (EU3P3ID6QAQ5R)
API Gateway: REST API endpoint (https://inr0anapu2.execute-api.us-east-1.amazonaws.com)
Region: us-east-1

Challenges we ran into

1. Agent Response Quality

Problem: Initially, all 4 agents gave identical generic responses regardless of their role.

Cause: Single AgentCore agent couldn't distinguish between different agent types from system prompt alone.

Solution: Built orchestration layer in Lambda that prepends role-specific instructions to each request. This allows one agent to behave as 4 specialized agents.

2. Session Memory & Context

Problem: Agents forgot previous messages in conversation, requiring users to repeat information.

Solution:

AgentCore's default memory feature stores conversation history
Generated unique session IDs per chat: action-1729697123-abc123
Passed same session ID throughout conversation
Memory persists until page refresh

3. Cost Optimization

Problem: Initially planned 4 separate AgentCore agents = 4x cost.

Solution: Single agent + Lambda role routing = $1/month vs $4/month (75% savings)

What we learned

AWS Bedrock AgentCore

How to deploy agents using agentcore launch CLI
Session-based memory configuration for conversation history
Model selection (Claude 3.5 Sonnet v2 optimal for legal reasoning)
Agent ARN/alias management for production deployment

Lambda + Bedrock Integration

bedrock-runtime.invoke_agent() event stream handling
IAM permissions: bedrock:InvokeAgent vs bedrock-runtime:InvokeAgent
Parsing streaming responses from AgentCore
Error handling for timeout, throttling, and model errors

Prompt Engineering for Legal AI

Role-based instructions must be very specific (generic = generic output)
Country/jurisdiction must be explicit in prompts
Format instructions (use bullets, bold, sections) improve readability
Examples in prompts increase output quality 10x

Frontend-Backend Integration

CORS configuration for cross-origin API calls
Session ID generation and management
Real-time feedback for async operations (file upload, API calls)
Error states and user feedback patterns

Cost Optimization

Bedrock pricing: $3 per 1M input tokens, $15 per 1M output tokens
AgentCore: ~$1/month per agent for memory + runtime
CloudFront: Cache-Control headers reduce origin requests 90%
Lambda: 512MB memory optimal for Bedrock API calls (faster = cheaper)

What's next for FreeLegal AI

Phase 1: Core Improvements (1 month)

[ ] Deploy 4 separate AgentCore agents (instead of orchestrated routing)
[ ] Add user authentication (AWS Cognito)
[ ] Store conversation history in DynamoDB
[ ] A/B test Claude 3 Opus vs Sonnet for legal accuracy

Phase 2: Advanced Features (3 months)

[ ] Contract Generation: Generate custom contracts from user requirements
[ ] Voice Input: Web Speech API for hands-free use
[ ] Multi-Language: Arabic, Spanish, French translations
[ ] Email Integration: Send demand letters directly from the app
[ ] Contract Templates: Library of freelancer-friendly contract templates
[ ] Payment Calculator: Estimate small claims court costs, collection agency fees

Phase 3: Enterprise & Monetization (6 months)

[ ] Lawyer Network: Connect users with vetted lawyers for complex cases
[ ] Team Features: Shared contracts, collaborative editing
[ ] Admin Dashboard: Usage analytics, A/B testing results
[ ] API Access: Let other freelance platforms integrate FreeLegal
[ ] Premium Tier: $29/month for unlimited conversations + priority support
[ ] White-Label: License to freelance platforms (Upwork, Fiverr, etc.)

Phase 4: Global Expansion (12 months)

[ ] 200+ Countries: Localized legal guidance for all jurisdictions
[ ] Industry-Specific: Agents for designers, developers, writers, consultants
[ ] Mobile Apps: iOS/Android native apps
[ ] Slack/Discord Bots: Integrate into team communication tools
[ ] Document Scanner: Mobile OCR for contract scanning
[ ] Legal AI Training: Fine-tune custom model on freelance contracts