LegalLens

LegalLens

Inspiration

Millions of people sign contracts they don't understand, leading to unexpected fees and unfavorable terms. I wanted to build an AI assistant that instantly identifies risky clauses and answers questions in plain language. AWS Bedrock's Nova LLM made this possible.

What it does

LegalLens analyzes contracts in three ways:

Risk Classification (🔴🟡🟢) - Flags dangerous clauses, cautions, and safe terms
Natural Language Q&A - Answer questions grounded in actual contract text
Semantic Search - Find relevant clauses using plain English

Impact: 50-page contract review → 40 seconds analysis + 3-second answers

How I built it

I engineered a RAG pipeline with AWS Bedrock:

Tech Stack:

Frontend: React 18 + Vite
Backend: Express.js + TypeScript
AI: Bedrock Nova-lite (classification & Q&A), Titan Embed v2 (vectors)
Storage: ChromaDB (vector database)

Pipeline: PDF → Text extraction → Chunking → Embedding → Classification → Q&A

Challenges I ran into

JSON Parsing - Nova wrapped responses in markdown. Fixed with regex stripping + fallback.
Performance - Sequential classification was slow. Solved with Promise.all() parallelization (5x speedup).
LLM Grounding - Answers were generic. Fixed by passing top-5 relevant chunks as context.
ChromaDB Stability - SDK conflicts. Switched to direct HTTP API calls.

Accomplishments I'm proud of

✅ Full end-to-end product (all 3 features working) ✅ Production-quality code (TypeScript strict, error handling) ✅ 40-second document processing, 3-second Q&A response ✅ Proper RAG implementation (no hallucinations) ✅ Comprehensive documentation