An ecommerce team was buried under repetitive questions in two languages. We built an assistant grounded in their own content that answers the routine ones and hands anything sensitive to a person.
Support agents spent most of their day answering the same questions about orders, returns, and delivery, split across Bangla and English. First replies were slow at peak times, and slow replies were costing the business both sales and goodwill.
Off the shelf chatbots did not handle Bangla well and had no safe way to touch real order data.
We built an assistant that reads their own help content and connects to their order and returns systems through a controlled interface. It replies fluently in Bangla and English, keeps to what it actually knows, and passes anything sensitive or uncertain to a human agent with the full context attached.
Every answer is traceable back to the source it came from, so the support team can see why the assistant said what it said and correct the underlying content when needed.
The core is a retrieval grounded approach: we index the client's content and order context, retrieve the relevant pieces for each question, and constrain the model to answer from them. We built an evaluation set in both languages and tuned prompts and retrieval against it, then added guardrails and a clean human handoff for edge cases.
"They understood that a support bot is only useful if the team can trust it. It handles the boring questions well and knows when to step back, which is exactly what we wanted."
Kevin Peterson, Head of Product