The Problem
Shopify merchants faced the same conversion challenge: generic product pages that showed every customer the same content regardless of their behavior, history, or intent. The product catalog was the same for someone who had purchased three times as for a first-time visitor. Browse patterns, past purchases, and real-time session behavior — all available, none of it used.
Product Sparks wanted to close that gap with an AI recommendation widget that any Shopify merchant could install and benefit from immediately — without hiring a developer, configuring a data pipeline, or waiting months for a model to learn their catalog. The widget needed to work with existing Shopify stores without customization, operate on serverless infrastructure to handle traffic spikes, and improve continuously as transaction data accumulated.
The Constraints
Zero-code installation was the product constraint. The addressable market — Shopify merchants — skews heavily toward business owners and operators with no engineering staff. A widget that required Shopify Liquid template editing or API configuration would eliminate most of the target market. Installation had to be a Shopify App Store install with no additional steps.
Cold start couldn’t be the excuse. A system that requires thousands of transactions before it produces useful recommendations is commercially useless for the median Shopify store. The model architecture had to produce reasonable recommendations from catalog metadata and session behavior from the first day of deployment.
Serverless or nothing. Shopify stores have unpredictable traffic spikes — product launches, flash sales, influencer mentions. An always-on server architecture would either over-provision (wasteful) or under-provision (outages during peak moments). AWS Lambda was the architectural requirement, not an option.
Our Approach
The recommendation engine combines a collaborative filtering model (trained on transaction history across the merchant’s catalog) with a session-based behavioral model (computing item similarity from real-time browse patterns within the current session). The two models are blended at serving time, weighted dynamically based on the amount of historical data available for that user — shifting toward collaborative filtering as purchase history accumulates, relying more heavily on session signals for new visitors.
The Shopify App installs as a standard app with no template modification required. The widget injects into the Shopify theme via a script tag, rendering recommendation carousels in configurable page locations (product page, cart page, home page) using the merchant’s existing theme styles.
Product catalog data syncs to DynamoDB at installation via the Shopify Product API. Session events (page views, add-to-cart, purchases) are captured client-side and processed through Lambda functions that update both the session model in real time and the collaborative filtering model in near-real time as transactions complete.
Redux manages client-side widget state; Next.js and React Router handle the widget’s micro-frontend rendering. Koa provides the backend API layer connecting the Shopify app with the recommendation engine.
The Outcome
- Personalized recommendations active on day one — even without purchase history, session behavioral signals drive relevant suggestions from the first interaction
- Installation completed by non-technical merchants in under 5 minutes — no developer required
- AWS Lambda handles peak traffic events without manual scaling intervention
- Recommendation accuracy improves continuously as transaction data accumulates per merchant
Client feedback: “Team delivered all tasks promptly, maintained clear communication, exceeded expectations.” — Melissa Bloggs, CEO, Product Sparks
Team
Engagement: 4 months, 3 engineers (1 AI/ML, 1 backend, 1 frontend).
Stack: Shopify, React, Redux, Next.js, React Router, Node.js, Koa, AWS Lambda, DynamoDB