Lexora — AI-Powered Contract Lifecycle Management (CLM)

Lexora is a full-stack contract lifecycle product: templates (NDA, vendor, consulting, employment, MSA), clause library, variable placeholders, rich editing, sharing, PDF/DOCX/Markdown/HTML export, and signature flows. Auth covers email/password, Google OAuth, verification and reset via Resend, JWT access tokens with rotating refresh, and RBAC. AI features use Google Gemini (with Ollama fallback) for chat grounded in the contract, risk-focused review, rewrite/explain/summarize, clause generation and suggestions, and streaming full-contract generation. Retrieval deliberately avoids vectors: a hierarchical tree is built from BlockNote structure (`contractTree.ts`); the model selects sections, pulls exact line-bounded sources, and server-side validation ensures citations quote real text—reducing hallucinated references compared to embedding-only RAG.

Timeline

Ongoing

Role

Full-Stack Developer

Team

Solo

Status

Completed

Source Code

Technology Stack

Next.js 14

React 18

TypeScript

Tailwind CSS

BlockNote

Node.js 20

Express 4

MongoDB

MongoosePassportGoogle GeminiOllamaResendJoiSwagger / OpenAPIpnpm workspaces

Key Features

Monorepo: `pnpm dev` runs Next.js (3000) and API (8080); OpenAPI docs at `/api-docs`

Contract workspace: editor, AI sidebar, variables, templates, outline, floating selection toolbar

PageIndex-style retrieval: heading tree, BFS candidate selection, line-numbered sources, citation validation

AI routes: chat, review, rewrite, explain, summarize, generate clause, suggest clauses

Platform: dashboard, notifications, image upload (Cloudinary or local), background jobs hook

Production paths: Docker for API, Render `render.yaml` with health checks

Key Learnings

Structuring CLM domains: templates, clauses, variables, signatures, and audit-friendly exports
Reasoning-based retrieval over document outline vs embedding similarity trade-offs
Monorepo discipline with shared types and env validation (Joi at startup)
Operating Gemini with retry/backoff and local-model fallback for dev resilience

Key Challenges

Grounding LLM answers in long legal documents without a vector database
Robust citation validation and fallbacks when the model drifts
RS256 key management and secure token rotation across client and server
Keeping editor state, tree index, and AI prompts consistent on every save

Impact & Results

Demonstrates production-shaped CLM plus novel vectorless RAG for contracts

Deployable backend pattern (Render) with documented secrets and health probes

Reusable reference for compliance-first legal tech UX

Future Enhancements

Dedicated production frontend URL and deep links from marketing site

E-sign integrations and jurisdiction-specific template packs

Automated regression tests for tree builder and citation validator

Org-level admin, billing, and audit logs for firm deployments