🧠 LegalScan – AI-Powered Proposal Classifier

LegalScan is a full-stack intelligent web app that classifies uploaded documents as formal proposals or non-proposals, using a hybrid of vector similarity (FAISS), heuristic rules, and LLM-based summarization. Built for legal professionals.

---

⚙️ Features

• 🔍 Proposal Classification
  Uses FAISS and keyword-based heuristics to classify documents as PROPOSAL, MAYBE_PROPOSAL, or NON_PROPOSAL.

• 📋 LLM Summarization
  Generates concise summaries of documents labeled as proposals.

• 📊 Rule Breakdown Display
  Explains scoring based on headers, keyword presence, structure, and submission terms.

• 🗂 Historical Dashboard
  Scanned documents are saved in a local database and accessible via a searchable UI.

• ☁️ Hybrid OCR Support
  Choose between local Tesseract OCR or AWS Textract for PDF/image parsing.

• 🛡️ Secure, Modular Design
  Environment-controlled secrets, RESTful architecture, SQLite or PostgreSQL-ready.

---

🧪 Example Use Case

Upload a government PDF. The app will:

1. Extract text (local or cloud).
2. Predict if it's a proposal.
3. Show a summary and rule score.
4. Log results to a dashboard for future reference.

---

🚀 Quick Start

1. Clone & Install

git clone https://github.com/your-username/legal-scan.git
cd legal-scan/backend
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows
pip install -r requirements.txt

2. Configure .env

FLASK_SECRET_KEY=your-secret-key
AWS_ACCESS_KEY_ID=your-aws-key
AWS_SECRET_ACCESS_KEY=your-aws-secret
AWS_REGION=us-east-1
S3_BUCKET_NAME=your-textract-bucket

3. Run the App

python app.py

Visit: http://localhost:5000

---

🧠 Classifier Logic

Each prediction is determined by:

• Vector similarity score (FAISS-trained embeddings)
• Rule-based boosters:
  • ✅ Header match (Request for Proposal, etc.)
  • ✅ Keyword density (budget, deliverables, etc.)
  • ✅ Structured sections (e.g., "Scope of Work", "Evaluation")
  • ✅ Submission language ("submit by", "deadline", etc.)

---

🗃 Tech Stack

Layer	Stack
Backend	Flask, SQLAlchemy
Frontend	Jinja2 + HTML/CSS
OCR	Tesseract (local) + AWS Textract (S3)
Classifier	FAISS + Regex Heuristics
Summary	LLM via Ollama/OpenAI/HuggingFace
Storage	SQLite (default)

---

📂 Directory Overview

backend/
├── app.py                  # Flask app entry
├── models.py               # SQLAlchemy model
├── templates/              # Jinja2 templates
├── static/                 # CSS & assets
├── faiss_utils.py          # Vector search classifier
├── ocr_utils.py            # Local Tesseract extraction
├── textract_s3_utils.py    # AWS Textract integration
├── llm_utils.py            # LLM summary agent
├── qualifier_rules.py      # Heuristic rule scoring

---

✅ Sample Output

Prediction: ✅ PROPOSAL (confidence: 0.87)
Matched Rules:
  - keywords: ✅
  - header: ✅
  - structure: ✅
  - submission: ❌

Summary:
This RFP outlines requirements for immigration legal services...

---

🧠 Why It Stands Out

• Combines NLP, LLMs, and classic AI for explainable document classification
• Modular, production-ready codebase suitable for legal tech, GovTech, and AI/ML portfolios
• UX designed around decision making, not just classification

---