InternAgent

InternAgent is an AI-powered communications engine built for the Internship Cell CET. It takes any internship URL or raw text, scrapes and parses the relevant details, and synthesizes them into ready-to-publish assets — an Instagram poster and a WhatsApp community caption — complete with a scannable QR code.

Built to solve a real workflow pain point: juggling multiple AI tools, copy-pasting between tabs, and formatting content from scratch every time a new internship drops.

---

Features

• Dual Input Modes — Scrape directly from a URL or paste raw text manually (useful for bot-protected or JS-rendered pages)
• Intelligent Scraping — Uses BeautifulSoup with a cloudscraper fallback for 403-blocked sites and an SSL fallback for misconfigured hosts
• SSRF Protection — Validates and resolves hostnames before any request is made; blocks private/loopback/multicast IPs
• AI Synthesis — Llama 3.1 70B (via NVIDIA NIM) condenses raw scraped content into structured poster copy and a WhatsApp caption
• Deterministic Splitting — Output is delimited by ===SPLIT===, parsed client-side into two distinct result panels
• QR Code Generation — Auto-generates a scannable QR for the application URL using qrcode.react
• Rate Limiting — Flask-Limiter enforces 5 requests/minute and 20 requests/hour per IP
• Decoupled Architecture — Static frontend on GitHub Pages, Flask backend on Render; API keys never touch the client

---

Tech Stack

Layer	Stack
Frontend	React 19, Vite 8, Tailwind CSS, Framer Motion, @shadergradient/react, qrcode.react
Backend	Python, Flask, Flask-CORS, Flask-Limiter, BeautifulSoup4, cloudscraper, Requests
AI	NVIDIA NIM — meta/llama-3.1-70b-instruct via OpenAI-compatible SDK
Deployment	GitHub Pages (frontend) + Render w/ Gunicorn (backend)

---

Local Setup

Prerequisites: Python 3.10+, Node.js 20+

1. Clone

git clone https://github.com/aswin-m-kumar/Intern_Agent.git
cd Intern_Agent

2. Backend

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

Create a .env file at the project root:

NVIDIA_API_KEY=your_api_key_here

Run the backend:

python app.py

Backend will be available at http://127.0.0.1:5000.

3. Frontend

cd frontend
npm install
npm run dev

The frontend auto-detects localhost and routes API calls to http://127.0.0.1:5000 in development.

---

Deployment

Frontend — GitHub Pages

The included GitHub Actions workflow (.github/workflows/deploy.yml) handles CI/CD automatically on every push to main. It builds the Vite app and deploys frontend/dist to GitHub Pages.

Note: vite.config.js sets base: '/Intern_Agent/' to match the repo name — update this if you fork under a different name.

Backend — Render

Deploy as a Web Service on Render with the start command:

gunicorn app:app

Gunicorn is pre-configured via gunicorn.conf.py with a 120-second timeout (necessary for LLM generation latency) and 2 workers.

Set NVIDIA_API_KEY as an environment variable in your Render service dashboard.

Point the frontend to your Render URL by setting a VITE_API_URL environment variable during the GitHub Actions build, or update the fallback URL directly in InternAgentForm.jsx.

---

Project Structure

Intern_Agent/
├── app.py                  # Flask app, CORS, rate limiting, /api/generate endpoint
├── internship_agent.py     # Scraping logic + NVIDIA NIM AI call
├── gunicorn.conf.py        # Gunicorn worker/timeout config
├── requirements.txt
├── .env                    # (gitignored) NVIDIA_API_KEY
└── frontend/
    ├── src/
    │   ├── App.jsx                        # Root layout, ShaderGradient background
    │   ├── components/
    │   │   ├── InternAgentForm.jsx        # Main form, state, API call, result panels
    │   │   └── LiquidGlassCard.jsx        # Glassmorphism card component
    │   ├── lib/utils.js                   # clsx + tailwind-merge helper
    │   └── index.css                      # Tailwind + custom scrollbar
    ├── vite.config.js
    └── tailwind.config.js

---

How It Works

1. User submits a URL or pastes raw text
2. Backend validates the URL (scheme, length, SSRF check via DNS resolution)
3. Page is fetched — with cloudscraper fallback on 403, SSL fallback on cert errors
4. BeautifulSoup strips scripts/styles/nav and extracts up to 20,000 characters of readable text
5. Text is sent to Llama 3.1 70B with a structured prompt instructing it to output poster content and a WhatsApp caption separated by ===SPLIT===
6. Frontend splits the response, renders both panels, and generates a QR code for the application URL