👻Introduction
AI Engineering student with a strong interest in research and development in Generative AI, particularly Diffusion-based image generation and Large Language Models. I’m building a solid foundation in Transformer architectures and optimization, supported by disciplined experimentation practices (clean experiment design, documentation, and reproducibility). I’m also hands-on with Python and deep learning ecosystems such as PyTorch / TensorFlow and Hugging Face. My current focus includes efficient fine-tuning and PEFT (LoRA / LoRA+ / other variants) for AI engineering under tight compute / VRAM constraints, as well as instruction tuning, Model evaluation and alignment. I also explore retrieval-based voice conversion (RVC) and multimodal pipelines. I am looking for a research environment with guidance and opportunities to turn ideas into testable, scalable real-world prototypes.
Education
- Sepuluh Nopember Institute of Technology - AI Engineering, 2024 - 2028, expected
- Semen Padang High School - Science, 2021 - 2024
Skill
🧊Experience
Bayucaraka UAV Research Team
🫧Achievement
Honorable Mention (Finalist)
AXION Kaggle Competition, 2025
1st Best Team
ISE! Academy Python Programming for Data Science Intermediate Level, Oct. 2024
🐋Top 5 Project
1) Waifu-Diffusion PEFT (LoRA) in Frieren Image Dataset
In this project, I fine-tune a lightweight LoRA adapter for the Waifu Diffusion text-to-image model ( hakurei/waifu-diffusion ) using the CyberHarem Frieren dataset, converts the dataset’s image-tag information into captions for training, trains the adapter with Diffusers’ train_text_to_image_lora.py workflow, exports the result as a compact pytorch_lora_weights.safetensors, and uploads it to the Hugging Face Hub so you can later attach it back to the base model with load_lora_weights() for inference.
Model License: CreativeML-OpenRAIL-M
2) MoeScraper
Python toolkit / library to help retrieve and collect image data from anime image fan art websites.
License: MIT
3) Implementation of Mixture of Low-Rank Adapter Experts (X-LoRA) Architecture in English-Indonesian Cross-Lingual Adaptation with Qwen2.5-0.5B
Attempting to implement the X-LoRA architecture in a Bilingual (English-Indonesian) task. The datasets used were CendolCollectionv2 for Indonesian and OpenOrca for English, both of which were pre-sampled to maximize results and save computation. Evaluation was conducted using BLEU, ROUGE-1, ROUGE-2, and ROUGE-L metrics.
Model License: MIT
4) Implementation of Fine-Grained Visual Categorization (FGVC) on The Quintessential Quintuplets Images Dataset using TransFG
Implementation of Fine-Grained Visual Categorization (FGVC) on The Quintessential Quintuplets Images Dataset using TransFG is a project that trains a fine-grained image classifier to distinguish between the five visually similar Nakano sisters (Ichika, Nino, Miku, Yotsuba, and Itsuki) using The Quintessential Quintuplets Images dataset. The pipeline fine-tunes a Vision Transformer with the TransFG idea of leveraging transformer attention to focus on the most discriminative local patches (“part/patch selection”), which is especially useful for FGVC where class differences are subtle and often concentrated in small visual cues rather than global shape. Using this TransFG-style setup, the trained model performs end-to-end inference to output the predicted sister label for a given image, achieving test loss = 0.4902 and test accuracy = 0.8433 (84.33%) on the held-out test split
License: MIT
5) Cross Lingual Web App: Waguri AI
Waguri AI is a bilingual chatbot web app built as a demonstration of fine-tuning Qwen2.5-0.5B using Mixture of LoRA Experts (X-LoRA) for the English–Indonesian pair. This web app is built using Next.js, Typescript, and Tailwind CSS for the frontend and FastAPI for the backend.
License: Apache 2.0
🤔Web App Idea for Final Project
Maybe... I want to build animanga LoRA library for Diffusion Model •⩊•
So people can search the LoRA adapter for diffusion model based on character name. Uhmm... for another variation, people can choose that the adapter was fine-tuned w/ Dreambooth or w/o Dreambooth.
Feature (target):
- Adapter search based on character name
- Dreambooth filter
- Adapter information
- Upvote or like button
Technology
- HTML, CSS, JavaScript
- Python
🌬️Dream Job
Become Haimiya-san's Husband AI Researcher
I want to be an AI Researcher cause I like to learn and build AI Architecture especiallly in Generative Model, and I think it's fun too :0
❄️ ご訪問ありがとうございます。 よい一日を！ (๑>•̀๑)