Commerce Recommendation Agent
What it takes to build an agent that gives personalized advice.
A good shoe fitting requires gathering personal context - running style, terrain preference, injury history, weekly mileage - then reasoning over a product catalog to find the right match. That makes it an ideal agent problem: multi-turn conversation, tool use against real data, and a recommendation that depends on who the user actually is. This project builds that agent end-to-end with Google ADK and On's real product catalog. Every architectural decision, tradeoff, and production lesson is documented as we go.
Tech Stack
Google ADK
Agent Development Kit - Google's framework for building multi-step, tool-using AI agents with built-in orchestration.
On.com Catalog
Real product data from On's running shoe catalog. No mock data - the agent works with actual products, specs, and pricing.
Web UI
A conversational interface where customers interact with the Commerce Recommendation Agent. Clean, responsive, and built for real-time agent responses.
Build Roadmap
Functional Requirements
PublishedDefine what the agent needs to do - user stories, conversation flows, edge cases, and what a good recommendation actually looks like.
Read the article →Product Catalog & RAG Pipeline
Coming SoonScrape On.com product data and build the retrieval pipeline - embeddings, vector store, and search interface the agent will use to find shoes.
Technical Design & Infrastructure
Coming SoonArchitecture decisions, component diagram, tech stack rationale, and infrastructure setup. The blueprint before any code.
Agent Scaffolding
Coming SoonSet up Google ADK, configure the agent loop, define tool interfaces, and get a basic conversational agent running locally.
Recommendation Engine
Coming SoonTeach the agent to gather context through multi-turn conversation and match customers to shoes based on running style, terrain, and preferences.
Web UI
Coming SoonBuild the conversational frontend, wire it to the agent backend, and make it responsive for real-time agent interactions.
Deployment & Production
Coming SoonShip the full system to production - hosting, monitoring, error handling, and the operational concerns that tutorials skip.
Evaluation & Iteration
Coming SoonAdd agent evaluation, test edge cases, measure recommendation quality, and iterate on the prompt architecture.