Context
In a large international retail group, a RAG proof of concept on Azure Prompt Flow was run in 2023. The agents now in production grew out of it.
Challenge
Moving from experimentation to uses that teams actually adopt, with access to AI services governed by policies.
What I did
My role: Tech Lead and Cloud & AI Architect since February 2024.
- A RAG agent that searches more than 6,000 IT contracts: documents flow from Google Drive to PostgreSQL and pgvector, and the contracts team, the only one with access, queries it from Google Chat.
- An n8n agent for incident managers outside business hours: the right contacts, resolution leads drawn from several hundred post-incident reviews, incident metrics for management.
- An agent for platform users: RAG over the documentation, viewing and creating Jira tickets through MCP.
- Governance: access to AI services governed by policies, models deployed on Microsoft Foundry and Gemini Enterprise Agent Platform, n8n agents in Docker on GKE.
- AI Champion for the Infrastructure department: weekly sessions with development teams on code assistants (Gemini Code Assist), and support for use cases.
Outcome
Three agents in production. Searching the contracts takes about a minute, instead of several hours by hand. Support requests from platform users fell by 30% over two years.
Stack
- RAG
- MCP
- n8n
- PostgreSQL
- pgvector
- Microsoft Foundry
- Gemini Enterprise Agent Platform
- GKE
- Docker