Benedict
Pepper.

Keywords: Deep Learning, PyTorch, Edge Deployment, Medical Ultrasound, ONNX Runtime

First author manuscript under review at Insyst (SINTA 3). Compares 3 lightweight encoders (ResNet-50, EfficientNet-B0, ConvNeXt V2 Atto) for fetal head circumference segmentation under simulated low-cost POCUS hardware conditions (physics-based speckle noise, acoustic shadows, blur). The best model (EfficientNet-B0) achieved 0.977 Dice at 322ms inference and 408MB RAM footprint, validated via 5-fold cross-validation (0.972 ± 0.003 Dice) and Wilcoxon signed-rank tests (p < 0.001) for 2-CPU/4GB-RAM edge devices.

Keywords: Health-Tech, Multimodal OCR, Gemini API, Patient Safety, PWA

An interdisciplinary research prototype developed for the National Student Creativity Program (PKM-KC). FitMate translates Mandarin herbal medicine (TCM) labels via Gemini multimodal OCR, cross-references a compiled 500-ingredient BPOM and SymMap database, and integrates a WhatsApp companion bot to prevent adverse drug interactions and improve patient safety.

Keywords: Applied Intelligence Systems, Python, XGBoost, MLP, Sensor Data

Published in Prosiding Seminar Nasional Universitas Ma Chung (Informatika & Sistem Informasi; Bahasa dan Seni; Farmasi) Vol. 6 No. 1 (2026), pp. 175–189, ISSN 2808-7771. Evaluated feature selection using ExtraTreesClassifier on 45,000 samples across 20 SIBI gesture classes and 16 physical sensor channels. Optimized via Optuna Bayesian search, XGBoost achieved 99.83% test accuracy and PyTorch MLP reached 99.08%, outperforming the 98.90% kNN baseline and resolving ambiguous overlapping gestures (Classes 16 & 17 F1-score: 0.997+).

Keywords: Deep Learning, EfficientNet, Transfer Learning, Progressive Resizing

A comparative study on training strategies for classifying Bacterial vs. Viral Pneumonia using Pediatric Chest X-Rays, published in Prosiding Seminar Nasional Universitas Ma Chung (Bahasa dan Seni / SENAM) 2026 and indexed in Garuda Kemdiktisaintek. The research demonstrated that fixed-size resolution training outperformed progressive resizing, achieving a final accuracy of 87.14% and an ROC AUC of 0.9431, optimizing computer-aided diagnostic reliability.

Keywords: UI/UX Design, System Usability Scale (SUS), Web Accessibility

Conducted a participatory UI/UX redesign of an e-commerce platform for special needs students' crafts. Through quantitative evaluation using the System Usability Scale, the intervention yielded a 25.4% improvement (from 57.81 to 72.50), elevating the platform's category from 'Marginal' to 'Good' and improving public visibility.

Keywords: Academic Communication, Linguistics, Content Analysis

A qualitative content analysis evaluating the formal vs. informal language usage in Artificial Intelligence course presentations. The study categorized syntactic structures and diction across multiple student groups, highlighting the critical role of formal linguistic norms in maintaining academic professionalism.