PromSketch-Standalone is a fully autonomous cloud telemetry runtime that supports high-throughput ingestion, sketch-based storage, PromQL-compatible querying, and visualization without relying on traditional monitoring systems. It features a parallel, event-driven ingestion pipeline capable of processing millions of time-series samples per second with significantly reduced latency and resource overhead. The system integrates a sketch-aware query and storage engine that performs real-time analytics directly over compressed metric summaries, providing high statistical accuracy while drastically reducing memory footprint.

Virtual Smart Assistant (VSA) is a personalized companion system designed to support students at Dian Nuswantoro University in both campus services and emotional well-being. It detects student emotions using facial recognition and a custom lightweight CNN for edge-based emotion tracking, and responds empathetically through an LLM-powered conversational engine trained on an Indonesian student-dialogue corpus. VSA features real-time noise suppression for clear voice interaction and is deployed on NVIDIA Jetson and Raspberry Pi devices, enabling adaptive, low-latency inference under limited compute and bandwidth conditions. The system aims to provide an accessible, responsive, and supportive virtual presence for students facing academic or personal challenges.

Low-end edge devices such as the Jetson Nano are constrained by limited computing power, memory, and graphics processing capacity, posing challenges for running complex models efficiently. To overcome these limitations, lightweight and fast models are essential to ensure real-time performance without lag, enabling timely and accurate emotional responses for users. In this research, we introduce the Custom Lightweight CNN Model (CLCM), a CNN-based face recognition model optimized for speed and efficiency while maintaining high accuracy. The results demonstrate that CLCM outperforms well-known pre-trained models, achieving up to 3× faster inference speed, with an average real-time detection time of 13 ms. Additionally, the model maximizes GPU utilization on the Jetson Nano, reaching 90% efficiency through CUDA integration, further enhancing performance without compromising device limitations.