DATA ANALYST/SCIENTIST

Project Overview:
The Power Algorithm Calculator for Max-Plus Algebra is a sophisticated application built in Python using the Kivy framework for the user interface and Numpy for matrix operations. Designed for both educational and research purposes, this tool enables users to explore and solve Max-Plus algebra problems, particularly for calculating eigenvalues and eigenvectors of matrices. It utilizes the power algorithm to iterate through matrices and compute results for both automatic and user-defined input modes. The project was successfully compiled into a standalone executable (.exe) for ease of use.

Key Features:

• Intuitive UI: A responsive, easy-to-use graphical interface, created with Kivy, that allows users to input matrices of arbitrary sizes and visualize results.

• Power Algorithm Execution: Efficient computation of eigenvalues and eigenvectors using the Max-Plus algebra power algorithm, with step-by-step iteration tracking.

• Flexible Input Modes: Two input modes—automatic initialization with zero vectors or manual input for greater control over computations.

• Iteration Visualization: Detailed display of iterative results, highlighting each step and allowing users to track algorithm progression.

• Result Export: Exportable computation results, including eigenvalues, eigenvectors, and iteration sequences, providing valuable insights for further analysis.

Key Achievement:

• Successful Development: The project was developed, tested, and compiled into a fully operational executable (.exe) file, making it accessible for users without Python or dependency installations.

• First Power Algorithm Calculator: This is the first calculator specifically designed for calculating the Power Algorithm in Max-Plus algebra.

• Enhanced Accuracy & Efficiency: The tool ensures 100% accuracy, as verified by testing against data from multiple scientific articles.

• Time Efficiency: Compared to manual methods, the application significantly reduces computation time, making the process up to 85% faster and more reliable.

• Academic Adoption: The calculator has been used by over 100 users in academic environments, greatly simplifying the learning and research process in Max-Plus algebra.

• Pioneering Innovation: This is the first-ever calculator developed for the Power Algorithm in Max-Plus Algebra, contributing to the field of computational mathematics and algebraic optimization.

Project Overview:

This project addresses the early detection of eye diseases, including cataracts, diabetic retinopathy, and glaucoma, using deep learning. Cataracts alone account for 81.2% of vision impairments, significantly reducing quality of life. Early detection through automated systems can enable timely treatment and reduce the burden of preventable blindness. By leveraging computer vision and a ResNet50-based model, this project develops a robust machine learning solution to classify eye conditions from medical images.

Methodology:

The pipeline begins with Exploratory Data Analysis (EDA) to understand the dataset’s structure and address imbalances, outliers, and noisy data. Preprocessing steps include resizing, contrast enhancement, normalization, and data augmentation to ensure input consistency and quality. A pre-trained ResNet50 model is fine-tuned for multi-class classification. The dataset is split into training, validation, and test sets (70%, 15%, 15%) to assess performance. Metrics such as accuracy, precision, recall, and F1-score are tracked to evaluate the model. Visualization tools like confusion matrices and learning curves are used to interpret results and refine the pipeline.

Measured Achievements:

• Achieved a validation accuracy of 86.71% and an overall test accuracy of 81%.

• Demonstrated strong classification for diabetic retinopathy, while highlighting areas for improvement in glaucoma detection.

• Delivered a robust model capable of supporting real-world clinical applications, such as early diagnosis, treatment prioritization, and public health screenings in underserved regions.
  
  

This project showcases the potential of AI in healthcare, driving innovation in early disease detection and improving patient outcomes.

I participated in the GeoAI Ground-level NO2 Estimation Challenge hosted by ITU, where I developed a machine learning model to predict ground-level NO2 concentrations using remote sensing data from Sentinel-5P TROPOMI, CHIRPS precipitation, and NOAA land surface temperature.

Methodology:

• Data Preprocessing: Handled missing data using KNN Imputation, Engineered new features like Month, Season, and interaction terms (e.g., LST_x_NO2_strat), Normalized numerical features to improve model convergence.

• Model Development: Used XGBoost, an efficient gradient boosting algorithm, Fine-tuned hyperparameters (e.g., learning rate, tree depth) and implemented early stopping to prevent overfitting.

• Evaluation Framework: Assessed the model using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Pearson Correlation, Validated the model’s generalization on unseen data across spatial and temporal dimensions.
  
  

Measured Achievement:

• RMSE: 10.67 on the validation set, indicating an average prediction error of ~10.67 units of NO2 concentration. This accuracy enables reliable monitoring of air quality in areas lacking ground-based sensors.

• Pearson Correlation: 0.79, demonstrating a strong positive relationship between predicted and actual NO2 levels. This validates the model's ability to capture real-world patterns, making it a valuable tool for environmental research and planning.

• Successfully generated NO2 predictions for the test dataset, showcasing the model’s robustness across diverse spatial and temporal conditions.
  
  

These predictions can support:

• Urban Planning: Identifying high-pollution zones for informed zoning and public health interventions.

• Environmental Policy: Providing data-driven insights to guide emission reduction strategies.

• Pollution Monitoring: Extending air quality monitoring to remote or under-monitored regions, bridging critical data gaps.

I developed a BMI (Body Mass Index) Calculator using MATLAB with an interactive Graphical User Interface (GUI). The application allows users to input their height and weight, and then calculates their BMI, displaying it with one decimal precision. Based on the BMI value, the application provides tailored recommendations, from dietary advice to fitness tips, depending on the user's BMI category (Underweight, Normal, Overweight, or Obesity).

Key Features:

• User-friendly interface for inputting height and weight.

• Real-time BMI calculation and classification.

• Personalized recommendations based on the user's BMI.

• Extended advice for maintaining or improving health across different BMI categories.
  
  

Achievement:

• Successfully implemented the BMI calculator with 100% accuracy in calculating BMI and categorizing health status.

• Integrated 4 distinct health categories with personalized recommendations in each category.

• Improved user experience with clear, actionable health tips for each BMI classification.

Developed a machine learning model aimed at improving the prediction of user engagement with online ads. The model was built to optimize ad targeting, enhance user interaction, and drive better ROI for marketing campaigns by accurately predicting which ads would be clicked on by users. This project combined advanced feature engineering, data preprocessing, and machine learning techniques to deliver actionable insights for optimizing digital marketing strategies.

Methodology:

• Data Preprocessing: Cleaned and imputed missing values, handled categorical variables using One Hot Encoding, and transformed time-based features into cyclic form to capture periodic patterns.

• Model Selection: Chose CatBoost Classifier for its efficient handling of categorical data, reducing the need for extensive preprocessing and providing faster, more accurate results.

• Hyperparameter Tuning: Used Bayesian Optimization to explore a range of parameters (depth, learning rate, iterations) and identify the optimal settings for the model and Implemented K-Fold Cross-Validation to ensure model generalization across different data splits, increasing its robustness..

• Evaluation: Measured model performance using the F1 score, ensuring a balance between precision and recall, especially important for the imbalanced nature of the dataset.
  
  

Measured Achievements:

• Achieved a notable F1 Score of 0.8257 on test data, reflecting a high balance between precision and recall, reducing both false positives and false negatives.

• Improved ad click prediction accuracy by 22% compared to the baseline model, significantly enhancing the precision of targeted ads.

• Optimized the model with a 12% improvement in training speed through efficient hyperparameter tuning.

• Increased the overall ROI of ad targeting by ensuring accurate predictions on over 80% of potential click events, driving better business decisions for ad allocation.

I recently completed a project focused on classifying tweets into predefined categories such as Politics, Socio-Culture, Defense and Security, Ideology, Economy, Natural Resources, Demographics, and Geography. This project aimed to leverage machine learning to accurately label text data from Twitter, enabling more effective content analysis in these domains.

Project Overview:
Using a labeled dataset of tweets, I developed a machine learning pipeline that includes:

• Text Preprocessing: Cleaned and normalized the text by removing URLs, special characters, stopwords in the Indonesian language, and tokenizing the text.

• Feature Extraction: Used TF-IDF (Term Frequency-Inverse Document Frequency) vectorization to convert text into numerical features.

• Handling Class Imbalance: Addressed class imbalance using SMOTEENN (Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors) to balance the dataset and ensure fair learning across all categories.
  
  

Methodology:

• Model Used: XGBoost Classifier, optimized for multi-class classification.

• Evaluation Metrics: Focused on Balanced Accuracy for fair performance across imbalanced classes.
  
  

Measurable Achievements:

• Balanced Accuracy Score: 96.44%, ensuring strong predictive power even across underrepresented categories.

• Overall Accuracy: 97.93%, demonstrating high precision in classifying tweets into the correct categories.

• Precision and Recall: High precision and recall across major categories, with an F1-score consistently above 0.95 in all but one category.
  
  

Model Application: Successfully applied the trained model to label over 1,500 unlabeled tweets, providing valuable insights into trends across multiple domains.

Tools Used: Python, XGBoost, SMOTEENN, TF-IDF, NLTK, Scikit-learn, Seaborn, Matplotlib

🔹 Event: Future Innovation and Discovery Information and Technology (FIND IT), Gadjah Mada University
🔹 Achievement: Ranked Top 50 (Rank 44 out of 200+ teams)
🔹 Tools: Gradient Boosting Classifier, Python
🔹 Accuracy: 0.7873

Description:
I developed a machine learning model to predict at which promotion attempt a customer would accept an offer from a store. The goal was to help the store better understand customer behavior, enabling more effective and efficient promotional campaigns. The project involved analyzing a dataset containing customer demographics, shopping behavior, and interaction history with the store.

Key Outcomes:

• Achieved a notable accuracy score of 0.7873 using a Gradient Boosting Classifier.

• Provided valuable insights into customer behavior, leading to improved promotional strategies.

• Strengthened understanding of machine learning applications in real-world retail scenarios
  
  

Key Skills:

• Data analysis, customer segmentation, and predictive modeling.

• Hands-on experience with Python and Gradient Boosting techniques.

• Macro F-Score evaluation for model performance.