In-depth Case Studies with Methodology, Metrics, and Impact

Objectives

• - Map digital inequality trends across Indonesia (2017-2024).
• - Diagnose the relationship between poverty, internet access, and ICT skills.
• - Support policy prioritization using hotspot identification and projections.

Dataset

• - Official BPS indicators: internet access, device ownership, ICT skills, poverty, HDI, IP-TIK.
• - Multi-year panel style data from 2017-2024, transformed into analysis-ready model.

Methodology

• - CRISP-DM pipeline: business understanding, data understanding, preparation, modeling, evaluation, deployment.
• - ETL, normalization, star-schema modeling, and DAX measures in Power BI.
• - Multi-level analytics framework: descriptive, diagnostic, predictive, prescriptive, exploratory.

Outcomes

• - National internet access trend increased to 89.8% in 2024.
• - Identified 11 provinces in digital exclusion hotspot category for intervention priority.
• - Produced 2025-2027 outlook for poverty and IP-TIK supporting policy planning.

Impact

• - Improved evidence-based discussion for digital inclusion policy.
• - Provided interactive scenario testing, not just static reporting.

Tech Stack

Power BI, DAX, Excel, CRISP-DM, Data Modeling

Objectives

• - Detect drowsiness through facial movement patterns accurately.
• - Provide real-time alert output on medium-to-low spec phones.
• - Keep inference efficient for practical use.

Dataset

• - 6 subjects, 22 videos, 720p, day/evening sessions.
• - Frame extraction generated 66,472 train frames, 16,835 validation, 9,328 test.
• - Labels: Alert (0) and Drowsy (1).

Methodology

• - Feature extraction with MediaPipe Face Mesh.
• - Derived EAR (Eye Aspect Ratio) and MAR (Mouth Aspect Ratio).
• - GRU training configuration: sequence length 24, epoch 49, Adam optimizer, early stopping, LR scheduler.

Outcomes

• - Test accuracy: 78%.
• - Class-level precision/recall/F1 around 0.78 for both classes.
• - Built end-to-end inference workflow from camera frame to drowsy/alert output.

Impact

• - Potential early warning mechanism to reduce fatigue-related accidents.
• - Demonstrated practical deployment path from model to user-facing app.

Tech Stack

Python, TensorFlow, MediaPipe, GRU, Android Studio, Kotlin, TFLite

Objectives

• - Classify text mood into six emotion classes.
• - Serve mood-aware movie recommendations in mobile flow.
• - Deploy API infrastructure in scalable cloud environment.

Dataset

• - TMDb metadata for movie recommendation candidate pool.
• - Emotion dataset from annotated Twitter text with 6 classes.

Methodology

• - Emotion classifier using LSTM architecture with embedding and softmax output.
• - Model conversion to TFLite for mobile compatibility.
• - Cloud deployment via Google Cloud Run and Cloud Storage support.

Outcomes

• - Model complexity: 7.6 million trainable parameters.
• - Achieved high training/validation performance in experimentation phase.
• - Delivered end-to-end app flow from mood input to recommendation page.

Impact

• - Created more personal entertainment recommendation experience.
• - Showcased cross-functional collaboration across ML, mobile, and cloud roles.

Tech Stack

TensorFlow, LSTM, Kotlin, Android Studio, Google Cloud Run, TFLite

Objectives

• - Predict liver disease risk using demographic and laboratory features.
• - Minimize misdiagnosis risk via stronger classification performance.
• - Benchmark ensemble performance against baseline models.

Dataset

• - Clinical dataset with demographic and lab variables such as bilirubin, albumin, enzymes, proteins.
• - Handled mild class imbalance using SMOTENC.

Methodology

• - Data cleaning, encoding, scaling, outlier handling rationale, and missing value treatment.
• - Stacking ensemble with base estimators (KNN, GaussianNB, LightGBM, SVC, XGBoost, Random Forest, CatBoost).
• - Meta estimator: Logistic Regression.

Outcomes

• - Accuracy: 82.85%.
• - AUC: 0.9128.
• - Recall: 79.23% and Precision: 85.19%.

Impact

• - Demonstrated improvement over several baseline models.
• - Useful as decision support companion to expert diagnosis.

Tech Stack

Python, Scikit-learn, CatBoost, XGBoost, LightGBM, SMOTENC

Objectives

• - Classify subreddit posts into suicidal vs non-suicidal categories.
• - Build reproducible NLP pipeline for large-scale social text.
• - Evaluate robustness with train-test split and cross-validation.

Dataset

• - 232,074 Reddit posts from related subreddits (balanced classes).
• - 80/20 train-test split with 10-fold cross-validation setup.

Methodology

• - Text preprocessing: lowercase, URL removal, mention/hashtag cleanup, digit cleanup, stopword handling.
• - TF-IDF vectorization with tuned settings (max_df, min_df, ngram range, sublinear TF).
• - Multinomial Naive Bayes model for binary text classification.

Outcomes

• - Accuracy: 92.9%.
• - ROC-AUC: 98.2%.
• - Reliable separation between suicidal and non-suicidal classes.

Impact

• - Illustrates how NLP can aid social good and mental-health support systems.
• - Strong benchmark project for high-volume text analytics.

Tech Stack

Python, NLTK, Scikit-learn, TF-IDF, Naive Bayes

Objectives

• - Automate scraping and ingestion from multiple local portals.
• - Provide searchable archive and export-ready dataset.
• - Improve usability through iterative product versions.

Dataset

• - News content pulled from 27+ online sources including local and national portals.
• - Structured archival data with source/date/category indexing.

Methodology

• - Backend architecture design with scraping modules and task progress management.
• - API endpoint development and data export workflow.
• - Frontend iteration from low-fidelity prototype to polished user-centered v2.0.

Outcomes

• - Reduced time for news collection and monitoring.
• - Centralized up-to-date local news archive for analysis.
• - Improved retrieval and documentation for long-term maintainability.

Impact

• - Increased operational efficiency for local statistical division.
• - Converted repetitive manual process into reusable digital workflow.

Tech Stack

Python, Flask, SQLite, BeautifulSoup, HTML, CSS, Pandas