Skip to content

Architecture

Overview

An overview of our system architecture.

alt text

Backend

Documentation for the Jenga backend can be found here

Frontend

Details about the user interface, web/mobile clients, and state management.

Deployment

Information on our CI/CD pipelines, infrastructure (like Kubernetes), and hosting.

AI

Our AI architecture is built around the Model Context Protocol (MCP), an open standard that allows AI models to safely interact with external tools and data.

Our main Quarkus application itself acts as the MCP Server. This is a powerful design choice because it means our AI agent has secure, direct, and high-performance access to all the project management data (tickets, users, etc.) from the inside. There is no slow or complex middle-man service.

Why We Chose This Architecture

  • Speed: By having the MCP Server built into the main Quarkus application, the AI can access project data instantly.
  • Simplicity: We don't have to maintain a separate AI service, database connections, or API. It's all part of one unified Quarkus application.
  • Standardization: Using the MCP standard means our architecture is modular. While we use Gemini today, we could (in theory) connect any MCP-compatible client to our server.

Core Components

  1. Quarkus (as the MCP Server): Our backend is more than just a web application; it's also a compliant Model Context Protocol (MCP) Server, thanks to the quarkus-mcp-server extension. This server's job is to define and expose "tools" (like getTicketDetails or findUser) that an AI model can use.

  2. Langchain: Inside our MCP Server, we use the Langchain framework. When a request comes in, Langchain is the "glue" that chains together the necessary steps. For example, it might:

    • Parse the user's request.
    • Call one of the tools (e.g., getTicketDetails).
    • Use the tool's result to build a new prompt.
    • Finally, send that prompt to the AI model.

  3. Gemini API: This is the external Large Language Model (LLM) that Langchain communicates with. We send it a well-structured prompt, and it provides the generative AI and reasoning capabilities to produce the final answer.

How It Works (The Flow)

  1. Request: A user's request is received by the Quarkus application.
  2. Process (MCP Server): The request is routed to the MCP Server module. Langchain takes over, calling internal functions (the "tools") to gather all necessary project data (e.g., ticket #123 details).
  3. Respond (Gemini): Langchain uses this data to build a final, context-rich prompt and sends it to the Gemini API. The AI's response is then sent back through the chain to the user.