Leboncoin deploys multi-agent LLM system to handle 60k customer service tickets monthly

Title

Leboncoin deploys multi-agent LLM system to handle 60k customer service tickets monthly

Content

Leboncoin, France's leading classifieds platform serving 28 million monthly visitors, deployed a three-agent LLM system that now handles 60,000 customer service tickets per month, automating 10% of tickets end-to-end with a stated 30% medium-term target. The system was designed and built with Hymaia to absorb a ticket volume that has grown to 200,000 per month, exceeding the capacity of the existing 30-model ML production fleet. The architecture uses a dispatcher that classifies each incoming ticket and routes it to one of three specialized agents. A generalist agent handles everyday questions via a Retrieval-Augmented Generation pipeline connected to the Zendesk FAQ, using FAISS for similarity search and an LLM for query reformulation. A buyer-transactions agent accesses Mondial Relay shipping data and payment information to resolve purchase-side queries. A seller-transactions agent mirrors that data for the supply side. When no agent reaches sufficient confidence, the ticket escalates to a human agent with the context already collected. The shift to multi-agent LLM orchestration addresses the structural challenge of running a marketplace at scale. The previous 30-model ML production fleet could not absorb 200,000 tickets per month, requiring a more sophisticated architecture with explicit RAG/DAG-over-fine-tuning approach for maintainability. Each component is independently testable and replaceable as models evolve. The system is built on Claude (third generation) accessed through Amazon Web Services, orchestrated with LangChain, containerised on Kubernetes and Docker, and integrated with Zendesk. Monitoring is split between Datadog for infrastructure and Langfuse for LLM-trace observability, with guardrails that moderate and constrain model output to enforce response quality. Architecture spans Claude (3rd gen) accessed via AWS, LangChain orchestration, Kubernetes and Docker containerization, Zendesk integration, and FAISS similarity search for the generalist agent. Monitoring uses Datadog for infrastructure and Langfuse for LLM-trace observability, with explicit guardrails for response quality enforcement. Operational scale at Leboncoin: 60,000 tickets per month processed through the AI system, with 10% end-to-end automation today and a 30% medium-term target. The team of about ten people operates 30 production models, six of which embed LLMs, treated as the lean-operations benchmark for GenAI inside a marketplace. Looking ahead, the deployment is rolling out at a measured pace — one new agent every three months — and the team is publishing the operational pattern internally so other GenAI projects across Leboncoin can reuse the same architecture, evaluation harness and hallucination-management playbook rather than relearn the productionisation path from scratch.