Accor - Plateforme d'agents de dev
An internal dev-agent platform, inside the AWS Accor perimeter.
- Status
- Internal project
- Year
- 2026
- Role
- AI Engineer - design & implementation (Accor engagement)
- Stack
- AWS Bedrock
- Agents
- MCP
- RAG
- pgvector
- Hono
- pg-boss
- Docker
- Nuxt
- TypeScript
Internal project: architecture and results shown without code or public link.
The context
Accor's web teams pile up mechanical tasks (migrations, dependency bumps, missing tests) and already have a headless agent and a model catalogue on Bedrock. The challenge: industrialize that into a governed platform, without any data leaving the AWS perimeter.
What I built
A multi-entry platform (Slack, GitLab issues, dashboard) wired to an orchestrator (Hono, pg-boss) that spins up ephemeral runners. Each runner runs an agent on Bedrock, leverages an MCP context server (design system, conventions, pgvector RAG) and opens a human-reviewed Merge Request. Costs and token budgets tracked to the cent.
Architecture
- 01RequestSlack, GitLab issue or dashboard, scoped then confirmed.
- 02OrchestratorQueue, priorities, budget and injected context.
- 03Runner + agentEphemeral container, agent on Bedrock, MCP server.
- 04Merge RequestTests passed, report attached, human review required.
- 05DashboardAcceptance rate, time saved and cost per MR.
Results & impact
- Inference 100% within the AWS Accor perimeter (Bedrock), no data leaving.
- A human always validates: agents open MRs, never merge.
- Per-task costs tracked and token budgets capped with alerting.
Skills applied
- Agents and orchestration
- RAG and MCP
- LLMOps and production
Got a project in mind?
An application to build, a team to strengthen or an AI product to launch: let's talk.