Qijin Xu

Qijin Xu · my-website-freedomjack.vercel.app

Enterprise RAG assistant for insurance support

CGI · 2023 – 2024 · Architect

About 90% of questions resolved without a person

An LLM-powered assistant that answers routine insurance questions on its own.

How the RAG assistant works A customer’s question goes to retrieval, which finds passages in the insurer’s own knowledge base. An LLM, orchestrated with LangChain, answers from those passages, and the answer goes back to the customer. Customer question several languages Retrieval RAG LLM LangChain Answer around the clock Insurer’s knowledge base Runs across Azure, AWS and GCP How the RAG assistant works A customer’s question goes to retrieval, which finds passages in the insurer’s own knowledge base. An LLM, orchestrated with LangChain, answers from those passages, and the answer goes back to the customer. Customer question several languages Retrieval RAG Insurer’s knowledge base LLM LangChain Answer around the clock Runs across Azure, AWS and GCP
The generic shape of the assistant, simplified. No client detail.
Read as steps
  1. A customer asks a routine question about a policy or a claim, in one of several languages.
  2. Retrieval finds the relevant passages in the insurer’s own knowledge base.
  3. An LLM, orchestrated with LangChain, answers from those passages.
  4. The answer goes back to the customer, around the clock. About 90% of questions are resolved without a person.
  5. The system runs across Azure, AWS and GCP.

Context

Insurance support teams spend much of their day on routine questions about policies and claims. The goal was an assistant that answers them from the insurer’s own knowledge base, in several languages, around the clock.

What I did

I architected and shipped the system: retrieval-augmented generation over insurance knowledge bases, built with LangChain across Azure, AWS and GCP.

Outcome

About 90% of questions resolved without a person, across a user base of 50,000+. 525% return on investment within 12 months. CGI Impact Award 2024.

Evidence

Both pages describe the product. They don’t name contributors or publish these figures.

Stack

Python, LangChain, LLMs, RAG, Azure, AWS, GCP

Qijin Xu is a Lead AI/ML Architect at CGI’s AI Center of Excellence in Toronto. He designs and ships production AI for banks and insurers: LLM agents, RAG systems, and data platforms.

All six projects →