← All projects

PROJECT / AI Agent Platform

Secure GPT: internal Generative AI platform

An internal generative AI platform for building AI assistants and products around sensitive company data.

  • AI Agent Platform
  • SecNumCloud
  • Fullstack Development
  • UX/UI Design

Overview

At a mutual insurer, many workflows involve highly sensitive data about our members. Public chatbots such as ChatGPT are not an option when that data cannot be sent to public services or processed outside France. Yet 99% of colleagues want a chatbot to help them work faster. Secure GPT was created to meet that need: an internal generative AI platform and API that let teams work with real data within the organization’s security requirements.

Development began in January 2025. Built on Gemini Enterprise Agent Platform and hosted on S3NS Trust Public Cloud (TPC) in an environment designed to meet SecNumCloud requirements, Secure GPT is currently available to colleagues in the IT department. It serves 2,000 users and offers around 100 agents and assistants in its application marketplace. More business teams are bringing forward use cases that can connect to the platform through its API.

My role

As a product-oriented fullstack developer within the AI Lab, I work at the intersection of product and engineering. I take the initiative to explore new AI capabilities and turn promising ones into practical improvements to the platform, its API and the workflows built on top of them.

Workflow

  1. Shape the solution: Teams bring forward a need. We clarify the users, data and workflow involved, then decide whether to use the platform directly, connect through the API, or create a tailored assistant.
  2. Prioritize and build: Business teams set priorities each week. Selected use cases become platform improvements, custom assistants or API integrations for internal products.
  3. Review and iterate: User feedback and security reviews help determine what to improve in the next cycle.

Key features

  • An API for internal products: Products can access the platform and its marketplace agents through the API, without having to build their own generative AI foundation.
  • Assistants grounded in company data: Member-facing applications can use real data to provide more relevant, personal responses. Grounding answers in trusted sources is intended to reduce hallucinations and help teams process cases faster.
  • Custom assistants: Teams can create assistants for focused tasks, such as answering questions from HR resources using retrieval-augmented generation (RAG), or helping draft email replies.
  • Company knowledge search: RAG over technical documentation can help employees find answers to questions such as how to create a delivery unit, where to deploy it, or which team to contact.
  • MCP integrations: Agents can connect to tools such as Jira and GitLab to help with ticket management and merge request workflows, reducing repetitive manual steps.

Architecture

Secure GPT exposes its capabilities through an API, so multiple products can use the same platform. The React application is one such client. A dedicated backend for each client controls which data it receives.

  • Client: The React and TanStack application is one of several clients. It authenticates through the internal gateway and uses an API key to access the platform.
  • API gateway: Apigee manages access to the platform API.
  • Client-specific backend: A Java and Spring Boot service (BFF) sits between each client and the platform, returning only the data that client needs.
  • Platform API: The core platform is built with Python and FastAPI and exposes generative AI capabilities to clients.