VICKY-GHOST AGENT

Undetectable AI agent using 4th-layer stealth and Gemini Vision for silent, zero-footprint data extraction.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for VICKY-GHOST AGENT

🛡️ Vicky-Ghost-Agent (v4.0): The Evolution of Silent Intelligence

🌟 Inspiration

As a Computer Science & Data Analytics student at IIT Patna, I’ve always been fascinated by the intersection of AI and Cybersecurity. The inspiration for the Vicky-Ghost-Agent came from a personal frustration: traditional web scrapers are fragile. They break whenever a website updates its CSS classes, and they are easily detected by modern anti-bot systems like those on X (Twitter) or LinkedIn.

I wanted to build a "Ghost" — an agent that doesn't just scrape code, but actually "sees" and "understands" the web like a human does. My goal was to create a tool that could navigate the secure web of 2026 silently, gather intelligence for my personal studio, Devil Labs, and leave no digital footprint behind.

📚 What I Learned

Building this project was a deep dive into advanced automation and Agentic AI. I learned:

  • Computer Vision > DOM Parsing: Using AI Vision to interpret data is far more robust than traditional scraping.

  • Network Fingerprinting: Learned how to spoof JA3 TLS handshakes to bypass server-level detection.

  • Session Persistence: Mastering Storage State and rolling cookie refresh logic is the only way to bypass modern login walls.

  • System Architecture: Designing a full-stack system that integrates a Python/FastAPI backend with a React 19 frontend while managing local system resources like DNS caches.

🛠️ How I Built It

The project is built on a High-Trust Stealth Architecture.

The Tech Stack

  • Backend: FastAPI (Python) for the core logic engine.
  • Frontend: React 19 + Tailwind CSS for the Command Center.

  • Intelligence: Gemini 2.5 Flash for vision-based data extraction and arrangement.

  • Stealth: Playwright (with Patchright) for browser automation.

Mathematical Stealth (Bezier Curves)

To mimic human mouse movements, I implemented cubic Bezier curves. Instead of a robotic straight line, the mouse follows a path defined by:

Where is the start, is the target, and are randomly generated control points that create a natural "jitter" and curve, making the bot indistinguishable from a human user.

The Workflow

  1. Identity Mapping: The agent selects the correct session from a Shadow Cookie Vault.

  2. Temporal Scrolling: It performs an infinite scroll until it visually identifies the target dates (e.g., Feb 8-9).

  3. Vision Extraction: Gemini "reads" a screenshot to extract fields like client_id, summary, and post_link.

  4. Intelligence Reporting: Results are compiled into a professional PDF.

  5. Self-Destruct: The system flushes DNS and shreds local files immediately after completion.

⚠️ Challenges I Faced

The biggest hurdle was the "Login Wall." Early versions of the agent would capture the "Join today" or "Sign in" screens on X.com because cookies were not properly injected. This resulted in empty extraction reports.

I solved this by:

  1. Implementing a Shadow Cookie Warmer that keeps sessions active by simulating background activity every hour.
  2. Developing a Multi-Key Rotator to handle 429 Resource Exhausted errors from the Gemini API, ensuring the intelligence layer never goes offline.
  3. Handling the Horizontal Space Error in PDF generation by switching to a more stable table-rendering API for complex data.

🚀 Conclusion

The Vicky-Ghost-Agent is more than a scraper; it's a testament to the power of combining AI with stealth engineering. It represents my journey as a developer at IIT Patna, pushing the boundaries of what an automated agent can achieve in a secure, modern digital landscape.

Built With

  • cookie-warmer
  • corsmiddleware
  • fastapi
  • fpdf2
  • gemini-2.5-flash
  • google-genai
  • ja3-tls-fingerprinting
  • patchright
  • playwright
  • pyinstaller
  • python
  • react-19
  • storagestate
  • tailwind-css
Share this project:

Updates