Environment
Agentic AI Platform — Veolia
Industrialising agentic AI at Veolia group scale: 50+ specialised agents organised in business kits, a fully custom group AI platform and a network of AI Leads trained across France and Morocco, for a community of 100 to 150 people.
- Client
- Veolia / Somei
- Role
- AI Platform Architect
- Period
- Mar. 2026 — Jan. 2027

The context
At Veolia, AI had so far been experimented with project by project, with no shared framework. My mission, as an external contractor: bring the group to scale by industrialising, governing and securing agentic AI for a broad, cross-functional community (developers, Product Owners, QA, UI/UX, agile project management), 100 to 150 people across France and Morocco.
Business kits, 50+ agents
At the core: complete AI kits, one per business function (dev, solution architect, product…). The platform now runs more than 50 highly specialised agents for the company, built on default company knowledge and knowledge bases split per kit, with a repo-first approach.
I designed full compliance for the kits and for artefact generation. Every agent follows a shared semantic convention, so each project gets an identical semantic architecture, with traceable, evolving artefacts (versions, snapshots…).
Agentic everywhere, from laptop to cloud
- Fully agentic cross-kit workflows: agents from several business functions work together on a single end-to-end flow.
- Platform-agnostic execution: the same agents run in VS Code, Claude Code, GitHub Copilot, Cursor or in the cloud.
- Headless agents running in isolated Azure environments (Azure Functions), governed through Vertex AI.
A fully custom group AI platform
The design authority platform has grown into the group’s AI platform, built entirely in-house. It lets teams:
- browse and run the kits, and update them remotely;
- see every project and, for each one, the AI in use through telemetry (FinOps included);
- launch cloud missions and fleet-wide campaigns, for instance bringing a whole set of repositories into architecture or security compliance.
Adoption & leadership
Tooling alone is not enough: a large part of the mission is adoption. I created and lead a network of AI Leads per chapter, each responsible for AI within their team. I train them, and they then spread the practices across their chapters. The programme is backed by talks and a variety of formats for every chapter, with on-site visits in France and Morocco to support adoption.
For professional confidentiality specific to this strategic project, certain details (precise architectures, telemetry, internal mechanics) are deliberately not covered here.
What to remember
A complete, custom-built, governed and measured agentic AI platform, from a single agent up to group-wide compliance campaigns. A network of AI Leads drives its adoption across 100 to 150 people in two countries.
Technologies
- Agentic AI
- Multi-agent orchestration
- GitHub Copilot
- Claude Code
- Cursor
- Azure
- Azure Functions
- Azure Container Apps
- Vertex AI
- Azure DevOps
- Terraform (IaC)
- OTel
- BigQuery & Looker
- Cosmos DB
- Azure Service Bus
- FinOps