Automatically generated narration. Code examples are available in the article.
From a chat request to an order
This project explored letting staff operate an E-Commerce management system through natural language. A request such as “Customer A bought three units of product B; create an order” gives the agent a task to complete through the existing backend APIs.
The agent uses LangChain and LangGraph with a ReAct flow to choose and call predefined tools. Those tools look up products and customers, check stock, and create orders.
My role
As Team Lead in a four-person team, I handled agent design, FastAPI backend implementation, tool definitions, self-hosted LLM integration, and AWS setup. I also built a simple JavaScript chat UI that shows results and confirmation messages.
The backend provides chat and agent-execution APIs, calls the existing E-Commerce API through tools, and works with MongoDB. For an order request, the agent can look up the product and customer, check the quantity, and call the order-creation API.
Working with self-hosted models
I evaluated self-hosted LLMs to explore a setup that did not depend on an external model API. I compared answer quality, tool-selection stability, response speed, infrastructure load, and operating cost.
The work made tool design particularly important: API contracts, input validation, confirmation before execution, and error handling all affect whether the agent can carry out a useful action.
Role: Team Lead · Team: 4 · Period: July–October 2025.