Conexia — Multi-tenant RAG platform for the Pacific
I built the multi-tenant AI chatbot platform that serves Conexia's retail, insurance, and services clients — from Tahiti.
- Kubernetes
- Scaleway
- Terraform
- Helm
- ArgoCD
- GitLab CI/CD
- Prometheus
- Thanos
- RAG
- LLM routing
- Vector DB
- Web scraping
Context
Conexia is an agency based in Tahiti that wanted to industrialise an AI chatbot offering for its Pacific clients: retail, insurance, services. The challenge wasn’t proving an LLM could answer a question — it was running a multi-tenant product, operable by a small team from an atoll, with reasonable latency despite the distance to European datacenters.
Three constraints shaped everything: cleanly isolating tenants (each client has its own data, sources, and secrets), reducing the cost of onboarding a new client to something measured in hours rather than sprints, and keeping the operational bar achievable for a small team.
My role
CTO. I owned the end-to-end design of the platform, the infrastructure choices, the agent code, the delivery pipelines, observability, and the definition of the client onboarding contract.
What I built
- RAG agents with semantic search over vector databases, routing across multiple LLMs based on cost and request criticality
- Rich conversational UI on the client side: interactive components, not a plain text stream
- Parallel web scraping pipelines that feed the vector databases from each tenant’s business sources
- Orchestration of 100+ workflows powering the agents (collection, normalisation, indexing, quality control, business triggers)
- Omnichannel connectors: web, Gmail, Messenger, all abstracted behind a common layer
- Scaleway infrastructure: Kubernetes clusters, Terraform for everything that ships, Helm + ArgoCD for GitOps, GitLab CI/CD with strict per-tenant secret isolation
- Observability: Prometheus + Thanos for long-term retention, real-time alerting on critical SLOs
- Client onboarding reduced to a single config file: zero custom code to add a tenant, which changes the agency’s economic model
The hard part: enforcing multi-tenant isolation cleanly at every layer (vector DB, secrets, observability, LLM quotas) without duplicating the platform per client.
Results
The platform is in production and continues to scale. Onboarding that used to require custom dev now happens through a config file, which changes the gross margin per client. Operations stay manageable for a small team thanks to GitOps and the observability stack.