Cenotium

The AI browser that instantly deploys secure, autonomous agents to book, pay, and execute your digital tasks—powered by robust data sharing and dependable protection for every step.

Comment

Inspiration

The modern internet is not built for AI agents. It was designed for human interaction—clicking, searching, reading, and manually executing tasks. However, with the rise of autonomous AI systems, a new paradigm is necessary—one where AI agents can interact, navigate, and execute tasks independently across the web.

We saw a fundamental gap:
-Websites are not built for agents – Most web pages lack structured APIs, making them inaccessible to AI.
-Autonomous AI requires a new interaction model – AI agents need a structured web representation to engage with content.
-Security & trust are critical – Autonomous systems need verifiable, secure, and efficient execution to operate safely.

This inspired us to redefine the internet—not by replacing it, but by overlaying a new structured schema that enables AI agents to navigate the web autonomously, fundamentally making us a New Age Browser.

🛠️ What We Built

A Browser for AI Agents

Cenotium is an Agentic Internet Browser, allowing AI agents to interact with and automate web tasks using a structured approach.

A Web Schema Transformer

  • Extracts website HTML, images, and UI components
  • Uses Vision, Grounding, and Action models to structure pages for AI interaction
  • Stores structured data in a Supabase database for efficiency and future page recall

An AI-Powered Agent Manager

  • Uses LLM Compiler to fine-tune user prompts using RAG from user preference knowledge graph
  • Divides tasks into single-action steps using LLMs
  • Delegates execution to autonomous AI agents

A Network of AI Agents

  • Perplexity Search Agent → Queries real-time information from the web
  • Twilio Calling Agent → Places calls & sends messages autonomously
  • Browser Activation Agent → Utilized E2B.dev containerized environments with OS Atlas grounding and LLM planning to interact with web elements, executing tasks via a Poetry-managed Python stack with PyQt5, Flask, OpenCV, and FFmpeg integration.

Security & Trust Layer

  • Global & Local Trust Core → Implements EigenTrust + PageRank for AI decision-making
  • Fernet Symmetric Encryption → Ensures secure agent communication
  • Persistent Storage (Redis & Supabase) → Maintains execution history & user preferences
  • Custom Inter-Agent Communication Protocol (RabbitMQ/Kafka) → Handles message security, encryption, and authentication

This architecture allows any new AI agent to be integrated, making Cenotium a scalable, future-proof AI ecosystem for web automation.

💡 What We Learned

  • The Web is Not AI-Ready – Most websites hide UI elements in React/Angular front-ends, making AI interaction difficult.
  • Planning & Execution Require Dynamic Replanning – Tasks must be broken into single-action steps and replanned on failure for seamless execution.
  • Security is Paramount – AI agents must operate with encryption, trust mechanisms, and secure communication to prevent malicious actions.
  • Agent Collaboration is Key – By allowing agents to call other agents, we built a self-improving multi-agent system.
  • Lack of Web Search Optimization Dataset - There is no structured dataset optimized for AI-driven web search and interaction, forcing agents to learn web navigation heuristically rather than relying on standardized training data.

🚧 Challenges We Faced

  • Building an Agent-Readable Web Schema → Converting unstructured HTML into structured agent-compatible schemas was a major challenge. We solved this using Vision + Grounding + Action models.

  • Efficient Prompt Optimization → Ensuring that the LLM Compiler created the most contextually aware user queries required RAG-based knowledge retrieval and continuous prompt tuning.

  • Security & Trust → We had to implement EigenTrust-style reputation ranking, Fernet encryption, and a custom inter-agent security protocol to ensure safe & fair agent execution.

  • Scalability → The system had to support infinite AI agents, allowing any new LangChain tool to become an agent. We optimized this using a modular, distributed execution model.

🛠️ Built With

Languages & Frameworks

  • Python → Backend development
  • JavaScript (Next.js, Tailwind, React & Streamlit) → Frontend & UI

AI Models & NLP

  • ChatGPT-4o, DeepSeek R1, Claude 3.5 Sonnet → Task decomposition, planning, execution
  • OS-Atlas deployed on HuggingFace → Bounding box detection for UI elements
  • OpenAI 4o-mini → Web image captioning
  • AWS Neptune → Knowledge graph Generation
  • Gemini Flash → Running of the Grounding model

Data & Storage

  • Supabase → Stores web schema & bounding boxes
  • Redis → Caches agent execution history

Security & Trust

  • Fernet Encryption + HMAC-SHA256 → Secure AI communications
  • EigenTrust + PageRank → AI agent reputation scoring

Agents & Tools

  • LangChain → AI agents built as modular tools
  • Selenium → Browser automation for web scraping
  • Twilio API → Automated phone calling & messaging
  • Perplexity AI API → Real-time web search agent
  • LLMCompiler → Assist Agent Manager to control and call agents

📸 Project Media

Agentic Internet

Cenotium is just the beginning—a paradigm shift redefining how AI interacts with the web. We are building a new internet, not for humans—but for agents. 🌎 🤖

Built With

Share this project:

Updates