Implementation of Parameter-Efficient Fine-Tuning for diffusion models

This project aims to develop a reliable and safe Indonesian-language medical chatbot specifically focused on digestive system health. We analyzed and fine-tuned 7 open-source Large Language Models (LLMs) using QLoRA (Quantized Low-Rank Adaptation) on a dataset of 21,376 doctor-patient Q&A pairs sourced from Alodokter.

Applied Parameter-Efficient Finetuning (LoRA+) to adapt Llama 3.2 1B for Indonesian conversational tasks. Achieved a 112% improvement in BLEU and a 21.7% increase in ROUGE-L compared to baseline models, demonstrating effective low-compute cross-lingual transfer learning despite domain shift effects.

FinSight is a comprehensive financial management platform designed to help Indonesian UMKM (Usaha Mikro, Kecil, dan Menengah) make better business decisions through smart financial analysis, cash flow predictions, and AI-based business recommendations.

This repository implements Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) models to forecast the foreign exchange rate of the Japanese Yen (JPY).

Designed and implemented a gastrointestinal polyp object detection system using deep learning. The project uses a clinically relevant gastroenteroscopic polyp dataset with annotated bounding boxes and benchmarks multiple state-of-the-art detection models (e.g., YOLOv8, YOLOv11, and transformer-based RF-DETR). Evaluated models using standard metrics such as IoU, mAP@50, mAP@50-95, precision, and recall to compare performance in real medical imaging conditions.

MiningChain is an agent-based decision support system that optimizes mining value chain operations using multi-agent simulation and operational data integration. It provides scenario-based recommendations to improve efficiency, mitigate risk, and support planning decisions across mining, logistics, and shipping activities.

Developed an end-to-end framework for automated analysis of Transcranial Doppler (TCD) signals to address data scarcity caused by poor signal quality. Implemented a Signal Quality Assessment (SQI) pipeline to stratify signals, followed by CycleGAN-based signal restoration to recover borderline-quality data. Trained and evaluated multiple 1D deep learning architectures, including CNN, ResNet, LSTM, Transformer, and a custom Attention CNN, for hemodynamic classification.