SentriX AI: Local-First Browser Level Data Leakage Prevention

Prevent sensitive data leakage to Large Language Models (LLMs) with 100% on-device processing.

---

🚀 Overview

As the adoption of Large Language Models (LLMs) accelerates, employees often inadvertently expose highly sensitive corporate data—such as API keys, Patient Health Information (PHI), proprietary code snippets, and customer PII—by pasting them into generic AI prompt boxes.

SentriX AI is a lightweight, strictly on-device browser extension that automatically monitors web text boxes (ChatGPT, Claude, Gemini, Copilot, etc.) and seamlessly intercepts document uploads (PDF, DOCX, Images). Before data ever leaves your computer, SentriX strictly pattern-matches and algorithms to detect vulnerabilities, offering users the ability to instantly Redact the information or heed a strict warning. All processing happens locally, maintaining a true zero-trust architecture.

---

✨ Key Features

1. Local-First Interception Architecture

   • Employs MutationObserver and Shadow DOM to dynamically inject interactive warning overlays without breaking the host application's UI.
   • Overrides fetch and XMLHttpRequest to catch attached files and documents pre-submission.

2. Multi-Format Document Parsing

   • PDF: Extracts text dynamically utilizing pdf.js and securely draws graphical redaction blocks over sensitive terms using pdf-lib.js.
   • Images (OCR): Implements local Tesseract.js WebAssembly workers to identify text within screenshots or scanned images, masking exact bounding boxes using HTML5 Canvas.
   • DOCX / TXT: unzips and alters XML directly inside the browser using jszip and mammoth.js principles to swap sensitive terms on the fly.

3. High-Precision Data Detection Pipeline

   • Comprehensive Regular Expressions leveraging precise capture groups to limit false positives.
   • Native detection for: PII (SSN, Phone, Email, PAN), Credentials (AWS/OpenAI/Stripe keys, Database URIs), Financials (Credit Cards), and Medical PHI (ICD-10, RxNorm patterns).

4. Context-Aware Smart Redaction

   • Doesn't just print [REDACTED]. It contextually masks data (e.g., swapping a real Stripe key for sk_test_your_stripe_key_here, generating dummy IP addresses, or leaving the database scheme intact while swapping the password).

---

🛠️ Technical Stack

• Framework: React + TypeScript via Vite (vite-plugin-web-extension)
• Background Processes: Chrome Extension Service Workers (MV3)
• DOM Manipulation: Custom Shadow DOM injection for absolute UI isolation
• Document Processing:
  • pdfjs-dist & pdf-lib (PDF manipulation)
  • tesseract.js (Optical Character Recognition)
  • jszip (DOCX parsing)

---

📦 How to Run / Develop Locally

1. Clone the repository:

   git clone https://github.com/Vedant03k/sentix-ai-llm.git
   cd sentix-ai-llm

2. Install dependencies:

   npm install

3. Build the extension:

   npm run build

4. Load into Browser (Chrome/Brave/Edge):

   • Navigate to chrome://extensions/
   • Turn on "Developer mode" in the top right.
   • Click "Load unpacked" and select the /dist directory generated by the build command.

---

🔒 Why SentriX AI Over the Competition?

Many Enterprise DLP (Data Loss Prevention) software solutions cost thousands per seat, require proxy routing, and send your data to third-party analysis servers. SentriX AI entirely eliminates the middleman. By analyzing entropy, context, and regex formats directly in the user's volatile RAM via WASM, no server infrastructure is required, meaning the extension itself mathematically cannot become a new vector for data breaches.

---

👨‍💻 The 7Seas Team Division

• Vedant Kale: PII detection
• Amrita Kumari: Extension Core, Shadow DOM architecture, and Platform routing.
• Pruthviraj Rajput: Complex Credentials Engine (Tokens, DB Strings, Credit Cards).
• Omkar Madkar: Internal File Override Engine, Multi-format Parsing (PDF/PNG/DOCX), and Medical Data Scanning.