Scalable Gemini-Based Code Review System: An Asynchronous Microservices Architecture for Learning Analytics

Manual inspection of students' GitHub commit histories to assess their programming activity presents significant scalability challenges when handling numerous repositories across a semester. This research developed and evaluated an AI-based commit analysis system that uses the Gemini language model to generate natural-language summaries of student programming activity, intended to support lecturers' review of GitHub-based coursework. The system was implemented using a prototyping methodology with an asynchronous, containerized architecture comprising a Vue.js front-end, a Laravel back-end, a PostgreSQL database, and a Redis-based job queue deployed in separate Docker containers. To keep the user-facing application responsive, AI analysis is processed asynchronously through a provider-abstraction layer accessing Gemini 2.5 Flash-Lite via OpenRouter, decoupling user interactions from computationally intensive model inference. System performance was evaluated using Apache JMeter on a 2-vCPU, 4-GB-RAM server, following the Performance Efficiency and Reliability characteristics of ISO/IEC 25010:2023. The back-end demonstrated a 0% error rate across 4,625 GitHub Feature requests. However, mean response time increased from 346.0 ms (10 users) to 6,743.8 ms (100 users), with throughput plateauing beyond 25 concurrent users, indicating a practical capacity threshold of 25–50 concurrent users. The AI Feedback path achieved a 100% success rate with a synchronous request-dispatch latency of 2.5–3.8 s across 90 requests; these are user-facing dispatch latencies and exclude LLM inference time. These results demonstrate the technical feasibility of the proposed asynchronous architecture for AI-assisted commit analysis at a moderate classroom scale; they do not constitute an evaluation of the accuracy, pedagogical usefulness, or learning-outcome impact of the generated feedback, which remain directions for future work. The primary limitation observed is performance degradation under high concurrency, which will require caching, database connection pooling, and vertical scaling for broader institutional deployment