This project focused on leveraging large language models (GPT-4 and Phi-3-mini) for automated Python program repair and hint generation in introductory programming scenarios. By incorporating advanced prompting strategies and Chain-of-Thought reasoning, the models were able to generate both actionable hints and code fixes. To ensure deployment efficiency, parameter-efficient fine-tuning was applied using LoRA, resulting in cost-effective inference. The approach achieved 88% average RPass and 15.23% average REdit with a compact 1.6 GB model, establishing a strong baseline and demonstrating the potential of LLMs in educational programming support systems.

This project involved building a multimodal monitoring system that integrates satellite imagery, Swissgrid electricity production data, and weather records to detect anomalies in nuclear power plant operations. The system was deployed using Docker, Kubeflow, MinIO, and Sentinel APIs, enabling scalable data processing and real-time analysis. Custom CNN-based models and an anomaly detection pipeline achieved 95% accuracy in identifying operational irregularities.

Developed an efficient object detection framework using knowledge distillation with two teacher–student model pairs. In the first setup, DINO + Faster R-CNN (Teacher) improved the accuracy of MobileNetV2 (Student) by ~10%. In the second setup, DETR with ResNet50 (Teacher) boosted the performance of DETR with ResNet18 (Student) by ~15%. This approach demonstrated how distillation can significantly enhance lightweight models, making them more suitable for resource-constrained deployments.

Indonesian Journal of Electrical Engineering and Computer Science