Amazon AI Shopping Assistant

Simple Interface of my agent

Inspiration

Navigating massive e-commerce catalogs often leads to choice overload and unhelpful search results. I wanted to build an intelligent conversational agent that goes beyond standard keyword searching—one that can understand user intent, filter catalog datasets dynamically using tool calling, and deliver tailored recommendations backed by real product metadata.

How I Built It

The application is built using Google Agent Development Kit (ADK) and Gemini 2.5 Flash on Google Cloud Platform:

Data Ingestion & Processing: Downloaded and structured product metadata from the McAuley Lab Amazon Reviews '23 dataset using custom Python parsing scripts.
Agent Architecture: Developed an LlmAgent using Google ADK, defining custom python function tools (get_products) that allow the model to query local catalog JSON files dynamically without inflating context tokens.
Frontend & Deployment: Built an interactive web interface using Streamlit and containerized the service for serverless deployment on Google Cloud Run.

Challenges I Faced

Payload & Latency Bottlenecks: Initially, passing raw dataset metadata directly into model tool responses caused high latency and slow rendering times. I resolved this by cleaning the raw Amazon schema into a lightweight, optimized structure with essential attributes (ASIN, title, rating, store, price).
Environment & Path Configurations: Troubleshooting Python package PATH configurations in Google Cloud Shell when setting up streamlit and google-adk binaries.

What I Learned

How to design function-calling tool interfaces with Google ADK to optimize token economy and reduce API costs.
Best practices for structuring RAG-style local catalog tools for LLM agent decision-making.
Deploying containerized Streamlit apps integrated with Vertex AI endpoints using Google Cloud infrastructure.