An LLM customer-service agent

I learned LangGraph on the job. I split the pipeline into stages with one job each: describe the customer's photo, rewrite the question as a search query, retrieve, then answer.

Customers mostly wrote in Mandarin and the knowledge index was in English, so I made it bilingual, and retrieval improved from matching in the same language. I added prompt-level constraints, including a short checklist the agent works through before it recommends a product.

For testing, I collected the cases customer-service staff sent back into a fixed set of real customer questions, and re-ran it at temperature 0 before every prompt change. A change went out only if it passed.

The pilot was strongest on product questions, with complex flows such as returns next in line. If I built it again I would add usage logging and automated grading from the first day.