Back to work

AI Document Review & Q&A Portal

RAG-powered legal document intelligence with e-signature workflows

Role

End-to-end (Frontend, RAG Pipeline, Ingestion Workers, Infrastructure)

Problem

Legal and business teams spend hours manually reviewing contracts and documents to answer questions, extract timelines, and identify risks—tasks that are repetitive but require high accuracy.

This platform transforms how teams interact with large document sets. Instead of searching PDFs manually, users ask natural-language questions and receive cited, sourced answers in real-time. Built as a full-stack RAG system, it handles document ingestion (PDF, Word, Excel, PPT), embedding generation, vector search with reranking, and LLM-powered summarization. Additional modules include timeline visualization, due diligence checklists, party/risk analysis, and a complete e-signature workflow with templating and audit trails.

The Challenge

Legal and M&A teams often work with hundreds of contracts, agreements, and diligence documents. Answering a single question—like 'What are the termination clauses?'—requires reading entire files, cross-referencing dates and parties, and summarizing findings. Traditional keyword search fails on synonyms and context. The goal was to build a system that could ingest any document format, understand semantic intent, retrieve relevant passages with high precision, and generate accurate, explainable answers with citations.

Technical Architecture

Frontend

  • Next.js (App Router)
  • React 19
  • TypeScript
  • TailwindCSS
  • shadcn/ui
  • React Email

RAG & AI

  • LangChain / LangGraph
  • Modal (vLLM hosting: Qwen2.5-32B)
  • BGE-M3 Embeddings + Reranker
  • @xenova/transformers (local fallback)
  • OpenAI-compatible endpoints

Data & Vector Search

  • Supabase (PostgreSQL + pgvector)
  • Full-text lexical search
  • Custom RPCs

Ingestion Workers

  • Modal Python workers
  • Marker PDF extraction
  • PyPDF2, textract, python-docx
  • OCR fallback

Infrastructure

  • Vercel (Frontend + API Routes + Cron)
  • Modal (LLM inference + ingestion)
  • Supabase Cloud / BYOD Supabase

Integrations

  • Clerk (Auth + Orgs)
  • Resend (Email)
  • pdfme (E-signatures)
  • Mixpanel (Analytics)
  • Sentry (Monitoring)

Key Features Built

  • Natural-language Q&A with streaming responses and inline source citations
  • Enhanced retrieval: BGE-M3 embeddings, reranking, lexical augmentation, and adaptive thresholds
  • Multi-format ingestion: PDF, Word, Excel, PowerPoint, TXT with OCR fallback
  • Timeline visualization: LLM-assisted extraction of dates, events, and parties from contracts
  • Due diligence checklists: Link documents to checklist items with completion tracking
  • Party and risk analysis: Extract entities, aliases, and assess risk via custom rules
  • Knowledge graph foundation: Items, links, and relationships
  • E-signature module: Templates (pdfme), guest signing, audit logs, final PDF storage
  • Multi-tenant SaaS with BYOD: Org-scoped data with encrypted customer Supabase credentials (AES-256-GCM)
  • Monitoring dashboards: Processing metrics, extraction timing, error tracking, cost summaries

Technical Challenges & Solutions

Challenge

Low retrieval precision on legal jargon and synonyms.

Solution

Switched to BGE-M3 embeddings (trained on domain-specific text) + reranker; improved precision by 40%.

Challenge

Large PDFs (1000+ pages) caused ingestion timeouts.

Solution

Batched chunking with progress callbacks; added job queue with retry logic.

Challenge

LLM hallucinations citing non-existent clauses.

Solution

Enforced strict context-only generation; added citation validation to ensure quoted text exists in source chunks.

Challenge

Multi-tenant data isolation with customer-owned DBs.

Solution

Built connection router with encrypted config storage (AES-256-GCM) and per-request DB client resolution.

←AI Real Estate Investment PlatformHealthcare Data Processing Platform→