DocuLens AI

Inspiration

The idea for DocuLens AI came from a real problem. A family member signed a 28-page apartment lease without reading it. Hidden on page 19 was a clause requiring 60-day move-out notice — they missed it and lost their security deposit.

Legal documents are everywhere: employment contracts, rental agreements, HR policies, eviction notices. But lawyers cost $300+/hour, and legal aid is overwhelmed. Most people sign without understanding what they're agreeing to.

We wanted to change that. Make legal understanding accessible to everyone — not just those who can afford a lawyer.

What it does

DocuLens AI is a legal document assistant that helps people understand contracts, policies, and notices without needing a legal background.

Upload any PDF — lease, employment contract, HR policy, eviction notice.

Get instant insights:

📋 Key clauses extracted and explained in plain English
⚠️ Risk alerts highlighting hidden penalties and unfair terms
🛡️ Your rights — what you should know, check, or ask about
💡 Actionable next steps

Smart features:

"Explain in simpler words" button for any clause
Ask questions about the document ("What's the late fee policy?")
Works 100% locally with no data leaving your computer

How we built it

Tech Stack:

Frontend: React with Tailwind CSS
Backend: FastAPI (Python)
RAG Pipeline: LangChain for document chunking + retrieval
Vector Database: ChromaDB for semantic search
LLM: Ollama with Mistral 7B (runs locally, zero API costs)
PDF Processing: PyPDFLoader

Architecture:

User uploads PDF → Backend saves and parses document
LangChain splits text into 1500-token chunks with 200-token overlap
Each chunk is embedded and stored in ChromaDB vector database
Mistral LLM extracts key clauses, risks, and generates analysis in JSON
Results displayed in React frontend with risk badges (🟢🟡🔴)
Users can ask follow-up questions via RAG retrieval

Why this stack:

FastAPI: Async support, automatic API docs, Python-native
ChromaDB: Local-first, no cloud dependencies
Ollama + Mistral: Completely free, private, runs on any laptop
LangChain: Simplifies complex RAG workflows

Challenges we ran into

1. Switching from OpenAI to local LLM
Initially we used GPT-4o-mini, but API costs added up during testing. We migrated to Ollama + Mistral — completely free, but required optimizing prompts for structured JSON output and handling slower inference (3-5 seconds per analysis).

2. JSON parsing failures
Mistral sometimes added markdown formatting or extra text before JSON. We built a clean_json_response() function that strips markdown and extracts only the JSON block.

3. PDF text extraction
Some PDFs were scanned images, not searchable text. We added error handling and clear messaging when a document can't be parsed.

4. Long document context limits
Mistral has a 8K token context. We truncate documents to 10,000 characters and use RAG to retrieve only relevant sections for Q&A, not the entire document.

5. Frontend-backend integration
Our teammate built the React frontend, and connecting it to FastAPI required careful CORS configuration and consistent JSON schemas between teams.

Accomplishments that we're proud of

✅ Working RAG system with local LLM
No cloud costs, no API keys, complete privacy. Anyone can run DocuLens AI on their laptop.

✅ Caught hidden risks in real documents
We tested on an actual employment contract and flagged a $25,000 penalty clause buried on page 4, plus a 3-year non-compete that banned any job within 100 miles.

✅ Plain English that actually works
Mistral generates explanations at 5th-grade reading level. Users don't need legal degrees to understand their rights.

✅ Clean, professional UI
Risk badges (🟢🟡🔴), collapsible clauses, one-click explanations, mobile-responsive design.

✅ Complete pipeline in 3 weeks
From zero to fully functional RAG application — PDF upload, vector search, LLM extraction, interactive Q&A.

What we learned

Technical lessons:

LLM prompting is an art. Small changes in prompt structure drastically impact JSON output quality.
ChromaDB is surprisingly easy to set up; local vector search is viable for documents under 100 pages.
Ollama makes local LLMs accessible — no GPU required for 7B models on a modern laptop.
RAG isn't just for chatbots; it's perfect for document analysis where you need to cite specific sections.

Teamwork lessons:

Clear API contracts (endpoints + JSON shapes) prevent frontend-backend conflicts.
Version control everything, including prompts — small prompt changes can break outputs silently.

Product lessons:

Users don't want "AI magic" — they want citations, quotes, and clear reasoning they can verify.
Legal tech has huge responsibility. We added a disclaimer: "Not legal advice. Consult an attorney for binding decisions."