Projects

A collection of my projects in AI, machine learning, and IoT. Check out more on GitHub.

2025

  • Exoplantropy — AI-Based Exoplanet Detection

    K. Ihsan, F. A. Fadilah, L. A. Djuanda, R. Yuniar, M. D. Hidayat, S. Muhammad NASA Space Apps Challenge 2025, Jakarta Global Nominee

    An AI-powered web application that detects exoplanets from space telescope data (Hubble). Our system increases the exoplanet identification rate from 1-in-10 to 7-in-10 by filtering false positives through machine learning classification. Developed during the 48-hour NASA Space Apps hackathon with mentorship from BRIN researchers and MIT AI experts.

  • CrackMap — Deep Learning Crack Detection

    Jun 2025 – Aug 2025 Universitas Negeri Semarang

    A deep learning-powered image analysis tool for detecting cracks in materials using computer vision. Provides a complete pipeline from image preprocessing (tiling, CLAHE enhancement, skeleton extraction, blur) to crack inference using a pretrained ResNet18 model. The heatmap output highlights crack regions — the brighter the area, the higher the model's confidence. A Streamlit web app is included for interactive demos with configurable tile division and minimum crack length parameters.

  • IHSG News Analyst — Predicting Stock Index Movement from News

    Mar 2025 – Aug 2025 Universitas Negeri Semarang

    An MLP-based predictive model that forecasts changes in the Indonesian Composite Index (IHSG) using multimodal features extracted from financial news and trading calendar patterns. Integrates semantic similarity from Sentence-BERT embeddings, financial sentiment signals from RoBERTa logits, and cyclical temporal feature encoding (dates, holidays, trading patterns). End-to-end pipeline from web scraping and preprocessing to modeling and Streamlit visualization.

  • IndoRoBERTa — Indonesian Financial Sentiment Classifier

    Jun 2025 – Aug 2025 MIT License

    Fine-tuned w11wo/indonesian-roberta-base-sentiment-classifier for Indonesian financial news sentiment classification (Positive / Neutral / Negative). Built because no financial sentiment model existed for the Indonesian market. Trained on intanm/indonesian-financial-sentiment-analysis augmented with Gemini-generated synthetic samples, GPT-2 prompt completions, and RoBERTa masked-token augmentation with class balancing (~23 k samples per class). Achieved 97.49% accuracy and F1-score (best at epoch 3 with early stopping). Trained on Google Colab T4 GPU. Published on Hugging Face.

  • Qwen2.5-0.5B-reasoning-lora — Reasoning-Enhanced Small LLM

    2025

    A LoRA-fine-tuned version of Qwen2.5-0.5B aimed at improving chain-of-thought reasoning capabilities in a compact language model. Demonstrates that lightweight parameter-efficient fine-tuning (LoRA) can enhance logical reasoning in sub-billion-parameter models, making advanced reasoning accessible on consumer hardware. Published on Hugging Face.

  • QLoRA Qwen2 0.5B — Quantized Fine-Tuning Experiment

    2025

    A QLoRA (Quantized Low-Rank Adaptation) fine-tuned Qwen2-0.5B model exploring efficient fine-tuning of small language models under extreme memory constraints. Uses 4-bit quantization combined with low-rank adapters to enable full fine-tuning on consumer GPUs while preserving model quality. Published on Hugging Face.

2024

  • Automated Plane Photography System with Real-Time Tracking

    Dec 2024 – Mar 2025

    An automated camera system that tracks and captures aircraft in real time using GPS-based flight data (ADS-B/API). Computes azimuth and elevation angles to dynamically control camera orientation via servo motors, enabling accurate visual tracking of passing aircraft. Built with embedded and IoT hardware (ESP32, ESP32-CAM, Raspberry Pi, SDR) and designed for future integration with ML-based aircraft recognition and cloud analytics.

  • InTeRRoR — Integrating LLM Output for Robotics Input

    Nov 2024 – Dec 2024

    An automated pipeline that translates natural language instructions into executable C++ code using a Large Language Model, then compiles and deploys the generated program directly to an Arduino-based robot in a single workflow. Enables high-level human commands to be converted into low-level embedded control logic without manual coding — from user instruction to robot behavior in one step.

Tech Stack

Python C++ PyTorch TensorFlow YOLO OpenCV Hugging Face Sentence-BERT RoBERTa Qwen LoRA / QLoRA Scikit-learn Streamlit Pandas Arduino ESP32 Raspberry Pi Serial / ADS-B