A comprehensive AI-powered data platform with hybrid cloud architecture designed for maximum security and flexibility when accessing private data sources.
agent/- IRA-Agent (Local Client) - Secure bridge to user's data sourcesserver/be/- IRA-Server Backend - Cloud control plane for user/team managementserver/fe/- IRA-Server Frontend - Web platform interface
This document summarizes the architectural plan for the IRA platform, its evolution from a secure database client to a comprehensive AI-native data platform, and the strategic approach for serving both individual and enterprise users.
The foundation of the platform is a hybrid model designed for maximum security and flexibility when accessing private data sources.
IRA-Server (Cloud Control Plane): The hosted web platform (backend + frontend). It manages users, teams, and connection metadata. It never stores or handles raw database credentials. Its primary security role is to act as a "Token Vending Machine."
IRA-Agent (Local Client): A lightweight, user-run application that acts as a secure bridge to the user's data sources (databases, APIs, etc.). It is the only component that ever connects directly to a user's private data.
Storage: When a user adds a data source, the credentials (e.g., DSN, API Key, OAuth Token) are stored directly in a secure vault like Google Secret Manager. The IRA-Server only saves a reference (resource name) to this secret.
Retrieval: To connect, the IRA-Agent authenticates with the IRA-Server, which provides the agent with a short-lived, narrowly-scoped permission token.
Connection: The agent uses this temporary token to fetch the credential directly from the secure vault and holds it only in-memory to establish the connection.
This model ensures sensitive credentials are never stored by the platform and are only ever transmitted between the secure vault and the user's machine. It is adaptable to any data source (databases, Confluence, Google Docs) and works seamlessly with complex networking like cloud DBs behind proxies, as the user's machine is responsible for the network path (via VPNs, SSH tunnels, etc.).
The core architecture is a launchpad for a much broader vision: transforming IRA into an intelligent data platform.
"Smart Company" Copilot: Use the agent as a secure ingestion pipeline for Retrieval-Augmented Generation (RAG). The agent can index private data, which the server uses to create vector embeddings, allowing an LLM to answer questions with relevant, private context.
Natural Language Interactivity: Enable users to ask questions in plain English. The IRA-Server interprets the request, and the agent executes the query and returns the data for the server to synthesize into an answer or visualization.
Ubiquitous Modality: The IRA-Server's API-centric design allows it to be the backend for various frontends, including Slack bots, Zoom/GMeet apps, or embeddable web chatbots.
A key challenge is balancing the high security of the per-user local agent with the usability needs of non-technical enterprise teams. A one-size-all approach is insufficient. The solution is a two-pronged strategy targeting distinct user personas.
Solution: The Local Agent.
Why: Perfect for technical users, connecting to localhost databases, and situations requiring maximum flexibility. These users are comfortable with a CLI and managing a local application.
Solution: The On-Premise Secure Gateway.
Why: Solves the usability problem for non-technical users. Instead of an agent on every user's laptop, the company's IT department deploys a single, hardened, persistent agent instance on a server inside their own secure network (e.g., AWS VPC, on-prem data center).
Key Features:
- Simplified UX: Non-technical users can interact via the web UI or Slack without ever needing to install or manage a local client.
- Preserved Security: The gateway runs within the company's trusted network perimeter, so sensitive data never leaves their environment.
- Auditing & Permissions: The gateway uses user impersonation. When it receives a query request from the IRA-Server, it connects to the database on behalf of the specific user who made the request. This preserves end-to-end auditing and allows the database to enforce its own fine-grained, user-specific permissions (e.g., row-level security).
This dual approach allows IRA to offer the best of both worlds: the ultimate security and flexibility of a local agent for individuals, and a seamless, scalable, and auditable solution for enterprise-wide adoption.
Each component has its own setup instructions:
- Agent Setup - Local client for secure data access (coming soon)
- Backend Setup - Cloud control plane (coming soon)
- Frontend Setup - Web platform interface (coming soon)
The platform currently provides AI models with secure database access through the MCP protocol, enabling natural language to SQL generation with proper authorization and security controls.