Architected and deployed an end-to-end AI microservices pipeline on a Linux VPS using Docker Compose, FastAPI, PostgreSQL, MinIO, and Vite. Developed an agentic web crawler powered by Browser Use, Playwright, and local LLMs (Qwen3 4B), enabling extraction of approximately 6,000 URLs per day while overcoming complex blockers such as popups, ads, and CAPTCHAs. Fine-tuned Indonesian large language models (Bakat-8B-Base and Qwen3-8B) using Continued Pre-Training (CPT), LoRA, and 4-bit quantization on a 92M-token corpus. Built a multimodal vision pipeline with Vision Transformers and RT-DETR for webpage content classification and UI element detection, achieving 99% accuracy and 0.99 F1-score.

Conducted technical reviews and debugging for 400-450 AI and Machine Learning projects monthly across six learning paths, including Generative AI, MLOps, Deep Learning, and Data Science. Evaluated advanced implementations such as custom training loops, Seq2Seq architectures, and Multi-Head Attention models while ensuring computational efficiency and preventing data leakage. Audited CI/CD pipelines, MLflow and DagsHub integrations, and monitoring systems using Prometheus and Grafana. Delivered structured feedback and mentoring insights to learners ranging from beginner to advanced levels.

Graduated with Distinction (Ranked 37th out of 616 participants) after completing 900+ hours of intensive AI training. Developed GoldSight, a GRU-based time series forecasting system for gold price prediction using a 60-day sliding window architecture, achieving an R² score of 0.9971 and MAPE of 0.62%. Performed hyperparameter optimization using Hyperopt and integrated financial news sentiment analysis with VADER to enhance market understanding. Deployed an interactive forecasting service using Streamlit for autoregressive multi-step predictions.

Designed predictive models for stunting risk detection in low birth weight and premature infants using XGBoost, Random Forest, and Logistic Regression on over 1,500 clinical records. Addressed severe class imbalance with SMOTE and class weighting, improving recall to approximately 0.89 and reducing false negatives. Conducted exploratory data analysis and feature selection to identify important clinical variables. Validated model robustness using 5-fold cross-validation, achieving a ROC-AUC score of 0.86.

Developed an end-to-end TensorFlow Extended (TFX) pipeline for sarcasm detection, including data ingestion, validation, feature engineering with TF Transform, model training, and automated evaluation using TensorFlow Model Analysis (TFMA). Applied hyperparameter optimization with KerasTuner and ensured reproducibility through containerized workflows. Selected among the top 597 participants from more than 60,000 applicants in the Expert level track.

Analyzed Indonesia’s air quality data using K-Means Clustering and Decision Tree Regression to identify key pollution patterns and trends. Clustered regions into five groups based on AQI and pollutant metrics (PM2.5, CO, NO₂, O₃). Determined optimal cluster count using Elbow Method and Silhouette Score. Built an interactive dashboard in Looker Studio to visualize insights and deliver actionable recommendations for air quality improvement.