Tetiana Afanasieva
Tetiana Afanasieva has a total of 2 co-authors. Here are the top co-authors by number of shared publications:
- Kateryna Babii
ELEKS Inc., Chicago, USA
1 shared publication - Serhii Onishchenko
Caterpillar Inc., Irving, Texas, USA
1 shared publication
Latest Publications
- Open Access
- 23 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 73479-73501
Web Application Performance Engineering Using Gemini AI
Web application performance is a primary determinant of user experience, scalability, and operational cost, yet optimization remains challenging in modern distributed architectures due to dynamic traffic patterns and heterogeneous diagnostic data, including execution logs, query plans, and workload traces. Conventional techniques such as load balancing, caching, and database tuning are effective but typically require extensive manual analysis and expert intervention [1]. This paper investigates the use of Google Gemini as a decision-support component for web application performance engineering. Gemini is not deployed as a runtime controller and does not perform real-time scheduling, resource allocation, or request routing; instead, it is applied during design and testing to analyze workload summaries, slow-query traces, and system logs and to generate actionable optimization recommendations [2]. We evaluate four advisory dimensions: (i) query optimization support, (ii) predictive caching recommendations, (iii) workload analysis for bottleneck identification, and (iv) behavior-aware caching strategies derived from aggregated workload-level user interaction patterns. Using a controlled benchmarking testbed representative of a modern Node.js–PostgreSQL–Redis web stack, we compare a baseline configuration against a Gemini-informed configuration in which model recommendations are implemented and validated by developers under identical workloads. Experimental results show that Gemini-informed optimizations reduce average response latency by up to 2–3×, increase throughput by up to 2× under high concurrency, and improve cache hit ratios to approximately 50% in read-dominant workloads through behavior-informed caching strategies, while maintaining comparable performance under low concurrency. These improvements are achieved without introducing the model into the production control loop. The results demonstrate that large language models can provide practical, explainable, and scalable guidance for web application performance optimization when positioned as advisory systems within established performance engineering workflows. This work presents an experimentally validated integration paradigm in which LLMs augment engineering decision making and complement infrastructure-level optimization techniques rather than replace them. A key distinguishing capability of this paradigm is cross-layer reasoning: by integrating SQL execution plans, application logs, and workload summaries into a unified advisory output, the LLM reduces engineering analysis time by an estimated 85–90% compared to manual performance debugging workflows.
- 23 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 73479-73501
Open AccessWeb Application Performance Engineering Using Gemini AI
Web application performance is a primary determinant of user experience, scalability, and operational cost, yet optimization remains challenging in modern distributed architectures due to dynamic traffic patterns and heterogeneous diagnostic data, including execution logs, query plans, and workload traces. Conventional techniques such as load balancing, caching, and database tuning are effective but typically require extensive manual analysis and expert intervention [1]. This paper investigates the use of Google Gemini as a decision-support component for web application performance engineering. Gemini is not deployed as a runtime controller and does not perform real-time scheduling, resource allocation, or request routing; instead, it is applied during design and testing to analyze workload summaries, slow-query traces, and system logs and to generate actionable optimization recommendations [2]. We evaluate four advisory dimensions: (i) query optimization support, (ii) predictive caching recommendations, (iii) workload analysis for bottleneck identification, and (iv) behavior-aware caching strategies derived from aggregated workload-level user interaction patterns. Using a controlled benchmarking testbed representative of a modern Node.js–PostgreSQL–Redis web stack, we compare a baseline configuration against a Gemini-informed configuration in which model recommendations are implemented and validated by developers under identical workloads. Experimental results show that Gemini-informed optimizations reduce average response latency by up to 2–3×, increase throughput by up to 2× under high concurrency, and improve cache hit ratios to approximately 50% in read-dominant workloads through behavior-informed caching strategies, while maintaining comparable performance under low concurrency. These improvements are achieved without introducing the model into the production control loop. The results demonstrate that large language models can provide practical, explainable, and scalable guidance for web application performance optimization when positioned as advisory systems within established performance engineering workflows. This work presents an experimentally validated integration paradigm in which LLMs augment engineering decision making and complement infrastructure-level optimization techniques rather than replace them. A key distinguishing capability of this paradigm is cross-layer reasoning: by integrating SQL execution plans, application logs, and workload summaries into a unified advisory output, the LLM reduces engineering analysis time by an estimated 85–90% compared to manual performance debugging workflows.