Architectural and engineering aspects of integrating cloud artificial intelligence services into an e-commerce web application

Abstract

Modern electronic commerce increasingly relies on artificial-intelligence capabilities such as visual search and product recommendation, yet for most mid-sized enterprises building and hosting the complete intelligence stack in-house remains impractical. A pragmatic alternative is to compose the required capability from managed cloud services; the engineering cost of such composition, however, is concentrated not in the models themselves but in the integration layer of the application. The paper investigates the architectural and engineering aspects of integrating five heterogeneous cloud services — a managed vector search index, a serverless inference platform, an object storage, a public model hub and a managed relational database — into a production web application for image-based dress search operated by a wedding fashion manufacturer whose catalog contains about twenty-nine thousand items and whose retail network spans more than three hundred sixty stores in seventy-five countries. A three-tier reference architecture is proposed in which every external service is isolated behind a dedicated adapter, authentication is encapsulated in a dedicated facade, and interchangeable embedding providers are implemented with the Strategy design pattern, enabling replacement of the vision-language model without modification of the business logic. The engineering decisions are formalized with analytical models: the availability of the synchronous composition of services, the success probability and worst-case latency overhead of a bounded retry scheme, and a break-even model for choosing between local inference and a per-call cloud interface. Experimental evaluation of the production deployment shows that external interactions dominate the request lifecycle: background removal through the serverless platform accounts for about three-fifths of the mean end-to-end response time of nearly four seconds, whereas the managed vector search contributes only about seventy milliseconds, or less than two percent. The local deployment of the domain-specific encoder surpasses the general-purpose cloud embedding interface in both retrieval quality and latency at the cost of several hundred megabytes of resident memory. During a four-week industrial trial, forty-five consultants executed more than eighteen hundred queries, and more than four-fifths of the queries received positive feedback, which confirms the practical viability of the proposed integration architecture.