NLP Application with OAuth Authentication and LLM-Powered Text Analysis

This project is a full-stack Natural Language Processing (NLP) web application built to explore how modern NLP systems—particularly LLM-backed features—can be integrated into production-style workflows with attention to reliability, evaluation, and safe engineering practices. Authenticated users can submit text for analysis, view structured outputs in a dashboard, and store analysis history for review.

The primary goal of this project was to learn end-to-end ML system design, including model integration, backend engineering, authentication, data persistence, and deployment.

---

Key Features

User Authentication

• Google OAuth 2.0 and Facebook login
• Cookie-based session management using cookie-session
• Sessions stored in signed, HTTP-only cookies
• Cookies configured with:
  • secure flag in production
  • sameSite: "none" for cross-site OAuth
• trust proxy enabled for Vercel deployments
• User profile display (name and avatar)

---

NLP Capabilities

The application supports multiple NLP tasks to compare local heuristics vs. LLM-based approaches:

Sentiment Analysis (Local Baseline)

• Local sentiment scoring using the sentiment library
• Outputs include score, comparative value, detected positive/negative words
• Used as a baseline to contrast with LLM-based reasoning

Text Summarization (LLM – Google Gemini)

• Uses Google Gemini to generate concise, paraphrased summaries
• Prompt constraints added to reduce verbosity and hallucination

Text Classification (LLM – Google Gemini)

• Classifies text into categories:
  • Bug report
  • Question
  • Complaint
  • Praise
  • Feature request
  • Other
• Designed to evaluate consistency across ambiguous inputs

Keyword & Entity Extraction (Heuristic)

• Custom keyword extraction utilities
• Simple named entity extraction for people, places, and organizations

Conversational Assistant (LLM – Google Gemini)

• Gemini-powered conversational interface
• Allows free-form interaction with structured prompt boundaries

---

Dashboard UI

• Tab-based navigation for each NLP feature
• Custom result cards per analysis type
• Chat-style UI for conversational outputs
• Keyword chips, entity columns, sentiment meters
• Expandable analysis history with delete functionality
• Fully responsive design

---

Storage & Data Handling

• NLP analysis results are currently stored in-memory for demonstration
• Each record includes:
  • id
  • inputText
  • operation
  • result (JSON)
  • createdAt
• Designed to support evaluation of model behavior over time
• Future work includes migrating analysis storage to Prisma ORM

---

Technology Stack

Frontend

• React
• React Router
• Custom dashboard UI and CSS
• Deployed on Vercel

Backend

• Node.js
• Express
• Passport.js (Google & Facebook strategies)
• Cookie-based session management using cookie-session
• Prisma ORM (planned for persistence)
• Serverless deployment on Vercel

NLP & AI

• sentiment (local sentiment baseline)
• Custom keyword and entity extraction utilities
• Google Gemini (@google/generative-ai) for summarization, classification, and chat

---

Security Considerations

• Explicit CORS allowlist with credential support
• Signed, HTTP-only cookies for session integrity
• Environment-based cookie security configuration
• Graceful error handling for LLM failures

Planned Hardening

• Per-user authorization checks
• Rate limiting and abuse detection
• Persistent session and analysis storage
• Structured logging and monitoring

---

Project Structure

root/
├── backend/
│   ├── server.js
│   ├── prisma/
│   │   └── schema.prisma
│   ├── package.json
│   └── .env
├── frontend/
│   ├── src/
│   ├── public/
│   ├── package.json
│   └── .env
└── README.txt

---

Engineering & ML Learnings

• Integrating LLMs into production-style APIs
• Prompt engineering for consistency
• Evaluating outputs through edge cases
• Failure-mode handling
• Separation of concerns
• Heuristic vs. LLM tradeoffs

Fallback behavior handles Gemini API failures gracefully.

---

Environment Variables

SESSION_SECRET=
GOOGLE_CLIENT_ID=
GOOGLE_CLIENT_SECRET=
GOOGLE_CALLBACK_URL=
FACEBOOK_APP_ID=
FACEBOOK_APP_SECRET=
FACEBOOK_CALLBACK_URL=
GEMINI_API_KEY=
FRONTEND_URL=

REACT_APP_API_BASE_URL=

---

Running Locally

cd backend
npm install
npm start

Runs at: http://localhost:3000

cd frontend
npm install
npm start

Runs at: http://localhost:3001

---

Deployment

• Backend deployed as serverless Express on Vercel
• Frontend deployed on Vercel
• Environment variables configured via Vercel dashboard

---

Future Improvements

• Automated evaluation metrics
• User feedback loops
• Confidence scoring
• Observability at scale