Abstract: Integrating large language model (LLM) APIs in an institutional web information system presents architectural challenges including service isolation, API fault tolerance, secured credential management, and cost-effective model routing. This paper presents service-oriented design patterns for embedding OpenAI GPT services within a Laravel-based system for community extension management, deployed at www.ceso.me. The architecture employs ten functional modules covering the full community extension lifecycle: user management with CSV bulk import and role-based access control (RBAC) across five user roles, activity management with unique entry code generation, content management system (CMS) and certification repository, and AI-powered feedback classification and automated reporting. The AI service layer uses a dual-model routing pattern: GPT-3.5-turbo (avg. latency: 0.8s, cost: P9.11/1K requests) handles dashboard insight synthesis, while GPT-4.1-mini (avg. latency: 1.4s, cost: P36.43/1K requests) performs granular sentiment classification with contextual reasoning, achieving 97% cost reduction compared to full GPT-4 deployment. Fault tolerance is implemented through exception-based fallback classification, achieving 98.7% API success rate with 1.3% graceful degradation during a 30-day observation period. Observed system performance metrics include average page load time of 1.2 seconds, database query time below 50ms, and successful handling of 500 concurrent users in load testing. Evaluation by 38 respondents using the ISO/IEC 25010 quality model yielded weighted mean scores of 4.82 (Functional Suitability), 4.55 (Performance Efficiency), 4.71 (Usability), 4.67 (Maintainability), and 4.75 (Decision Support Effectiveness). The design patterns and architecture presented are intended to be adaptable to institutional web systems seeking to integrate LLM services with production-grade reliability.