Projects List

An open-source and privacy-first browser extension designed to translate manga directly on your device without requiring logins or proxy servers. The system features a dual-pipeline architecture that includes a completely local mode utilizing YOLO26 for bubble detection, PaddleOCR for text extraction, and a custom line splitter feeding into WebGPU-accelerated Qwen3 via WebLLM for translation. It also offers a cloud alternative that securely routes annotated images directly to the Gemini API. Engineered for broad compatibility using custom regex parsing rules and thoroughly tested on MangaDex, the tool includes an opt-in telemetry system for continuous detection model retraining.

An implementation of a Res-U-Net architecture adapted for the DCASE 2025 Sound Event Separation task. This project modifies the music separation model to handle overlapping domestic sound events using a custom dataset of 10,000 samples. Built with PyTorch, the model specifically integrates Time-Frequency Sequence Attention (TF-SA) mechanisms to effectively capture long-range dependencies in magnitude spectrograms. The model achieves a Class-Aware SDR improvement (CA-SDRi) of 5.25 dB, significantly outperforming standard Res-U-Net baselines.

An automated end-to-end NLP pipeline designed to map urban issues and public sentiment in Surabaya using social media data (X/Twitter and Reddit). The system features a custom scraper, fine-tuned BERT models for sentiment analysis, and a custom Named Entity Recognition (NER) model specifically trained to extract local locations. Fully automated via GitHub Actions for weekly data collection and monthly topic model retraining, the analysis is visualized on a real-time Streamlit dashboard hosted on Hugging Face Spaces using Supabase for persistence.

As the sole developer, I engineered Action Unesa, a comprehensive competition management platform built from the ground up to automate and streamline all operational aspects for organizers and participants. This full-stack application features a robust admin panel with a real-time dashboard, an advanced team verification system with automated notifications, and powerful bulk actions for editing submissions and sending targeted announcements. To ensure data integrity, I implemented safety features like soft-deletes and a full admin audit log. For users, the platform offers a seamless experience with Google OAuth, intuitive team creation via invite codes, and a personalized dashboard. The entire project was built on a modern stack, utilizing Laravel for the backend, React with TypeScript and Inertia.js for the frontend, and an automated CI/CD pipeline with GitHub Actions for continuous testing and deployment.

This project is an AI-powered system designed to assist in critical search and rescue (SAR) missions. It analyzes thermal imagery from Unmanned Aerial Vehicles (UAVs) to locate missing persons quickly, especially in low-visibility conditions. The goal is to provide first responders with a tool to drastically reduce search times and increase the chances of a successful rescue. The system is built on a dataset of 4,874 annotated thermal images and utilizes two optimized object detection models: RT-DETRv2 and YOLOv12. These were chosen for their critical balance of accuracy and speed in time-sensitive SAR applications. Key features include high-accuracy detection of human heat signatures, real-time performance, and a specialized focus on SAR scenarios.

This project is a personal portfolio website designed with a striking Neobrutalist aesthetic, built to serve as someone's central online presence. Leveraging the power of React for dynamic interactivity and Tailwind CSS for utility-first styling, the site is structured as a clean, single-page application. It features a modern, minimalist interface with a high-contrast dark mode and a fully responsive layout that ensures a seamless experience on any device. A key highlight is the innovative use of Iframe previews, which allows visitors to directly interact with hosted web projects without leaving the portfolio, offering a clear understanding. This portofolio can easily being replicated using the tutorial in the github repository

RoastX is an innovative and entertaining web application designed to generate witty, sarcastic roasts of a user's X (formerly Twitter) profile using the power of Gemini AI. Built with PHP for optimal performance and a small application size, the project scrapes public data from the X clone, Nitter, using cURL and HTMLXPATH to analyze a user's profile. This data is then fed to the Gemini AI, which crafts a unique and humorous roast. The frontend is built with React JS and styled with Tailwind CSS, creating a modern, responsive, and aesthetically pleasing user interface. To enhance the user experience, the application incorporates Framer Motion for engaging and dynamic animations, making the roasting process not just a service but an interactive and fun experience.

This project is a clean and elegant landing page designed for "Semoga Lestari," a brand offering premium accommodations. The primary goal was to create a strong first impression that communicates luxury, comfort, and exclusivity. The user experience is crafted to be seamless and intuitive, ensuring that visitors can immediately grasp the brand's identity and offerings with a clear call-to-action.

Study-World is a dynamic and interactive e-learning platform built with the versatile Flask framework, designed to create an engaging space for learning and collaboration. For administrators, the platform offers a comprehensive dashboard to manage users, content, and view site-wide statistics. A key feature is the dynamic module system, allowing for the easy creation, modification, and removal of educational content, including quizzes that reward users with points. To foster a motivating learning environment, Study-World incorporates a points and leaderboard system, along with a comment section in each module to encourage user interaction and discussion.