Multi-Agent Customer Support System
Key Results
- › Master + Policy (RAG) + Ticket (tool-using) agents
- › Evaluated with 15 scenarios and an LLM-as-a-judge
- › Open source, with an 8-part write-up
- › Also rebuilt in LangGraph
What this is
I wanted to build a genuinely production-shaped customer support agent — not a toy demo — and show every part of it. So I built one end to end for a fictional UK transport company, “UKConnect,” and wrote it up as an 8-part series.
To be clear up front: this is a build project, not a client deployment. UKConnect isn’t real, so there are no real customers and no measured cost savings. What’s real is the system and the engineering.
The approach
A multi-agent system powered by Google Gemini, with two specialists: a Policy Agent that uses RAG to answer policy questions from company documents, and a Ticket Agent that calls tools to search, book, modify and cancel against a database. A Master Agent reads each message, routes it to the right specialist, and synthesises one reply. It’s served with FastAPI, with logging throughout.
What’s real about it
It’s a complete, working system, and it’s open source. I evaluated it with 15 end-to-end scenarios and an LLM-as-a-judge rather than brittle string matching, and I later rebuilt the whole thing in LangGraph to compare the two frameworks. The full walkthrough — architecture, data layer, RAG, agents, orchestration, testing — is in the series.
If you want a system like this built for your business, get in touch.