shopee AI

A e-commerce AI assistant that actually work.

Comment

Inspiration

We’ve all been there: stuck in a loop with a generic chatbot on AliExpress that gives canned answers and never actually solves the problem. We were tired of "AI Support" being a barrier instead of a bridge. We wanted to build a customer support agent for Shopee that doesn't just talk—it acts. Our goal was to create an AI that possesses the tribal knowledge of the community and the logistical power of a human agent.

What it does

Shopee AI is an intelligent, action-oriented support ecosystem designed to replace the frustrating experience of traditional e-commerce bots. It functions as a high-tier customer support representative that can:

  • Resolve Complex Queries: Uses a RAG (Retrieval-Augmented Generation) pipeline to pull answers from internal documentation and official guides.
  • Crowdsource Solutions: Taps into community chat history to find "real-world" fixes that aren't always in the official manual.
  • Bridge the Inbox: It has the authority to read incoming customer emails and draft/send context-aware replies, managing the entire support ticket lifecycle.

How we built it

The project is built on a robust AI stack designed for accuracy and speed:

  • The Brain: Powered by Gemini 3 Flash for fast, reasoning-heavy natural language processing.
  • The Memory (RAG): We utilized a vector database to index thousands of internal Shopee docs, seller guides, and community forum threads.
  • The Interface: A React-based chat widget integrated into the Shopee ecosystem.
  • Email Integration: Connected via Gmail/Outlook APIs, allowing the model to parse threads and generate responses using structured function calling.

Challenges we ran into

The biggest hurdle was the "Noise-to-Signal" ratio in community chats. Forum data is inherently messy—full of slang, typos, and occasionally incorrect advice. We had to build a custom pre-processing layer to rank community solutions based on helpfulness before feeding them into the RAG pipeline. Additionally, managing API key lifecycles and KES project configurations taught us hard lessons about production-level security and deployment stability.

Accomplishments that we're proud of

We are incredibly proud of our Action-Oriented Architecture. Most bots just point you to a link; Shopee AI can actually draft the email for your refund request or explain a complex shipping policy using data synthesized from three different internal sources simultaneously. The latency is impressively low, making the conversation feel human and fluid.

What we learned

We learned that Context is King. An AI model is only as good as the data it can reach. Integrating "unstructured" data like community chats provided a level of empathy and "human" insight that official documentation simply couldn't offer. We also gained a deep understanding of the importance of robust credential management in a live e-commerce environment.

What's next for Shopee AI

The next step is Proactive Support. We want Shopee AI to detect when a package is delayed before the customer even asks and automatically draft an apology email with a discount code. We also plan to expand the RAG to include video tutorial transcripts, allowing the bot to send timestamped video help directly to users.

Built With

Share this project:

Updates