bloggenx
genx
Twelve microservices, one platform, and every tenant sure it is the only one on it.
the idea
the problem
Most enterprise AI pilots die in the demo. Every team wires up its own model keys, its own retrieval, its own connectors, and none of it survives a second customer. The model call is not the expensive part. Everything around it is: routing, memory, ingestion, permissions, and the fact that none of it is shared.
the approach
GenX is a multitenant platform that treats all of that as infrastructure. One agent runtime, one retrieval stack, one connector layer, and tenants isolated above them. Teams say what they want; the platform handles the calls, the context and the plumbing.
what's done
Agent runtime with a tool loop, streaming responses, retries and per-tenant limits.
Hybrid RAG: dense plus lexical retrieval, reranked, over a shared ingestion pipeline.
Multi-provider routing through LiteLLM, so a tenant can switch models without a rewrite.
MCP connectors, conversation memory and data ingestion across all twelve services.
Deployed on Kubernetes, with isolation per tenant.
what's missing
A public demo. I built the platform for customers, not for browsing.
Self-serve onboarding. Right now a new tenant is a conversation, not a form.
Public benchmarks. The internal evals exist, but I cannot publish them yet.
how to try it
The repository is private. If you want a walkthrough, mail me at [email protected] and I will show you the architecture.
The surrounding pieces are public: github.com/Suhaas-code.