Work

Projects

Selected work in AI, machine learning, and computer vision.

• 01

  MARSHALL

  Learning module automation platform at Telkom Indonesia — used by 1,406 corporate employees. Built with GPT-4o and Llama 3.2; advanced RAG pipeline improved response factuality from 20% to 90%, reducing manual workload by 40% over six months.

  GPT-4o Llama 3.2 RAG FastAPI LLM

• 02

  Computer Vision Guided Autonomous Drone

  Object detection with YOLOv3s + TensorRT on Jetson Nano to track pickable objects and avoid obstacles. ROS-based odometry using barometer for velocity and altitude — achieving 90% pick accuracy in real-world testing. 1st Runner-up at FIRA Air SimulCup 2022 and KRTI Regional 2023.

  YOLOv3 TensorRT ROS Jetson Nano Computer Vision

• 03

  Face Recognition Attendance System

  Online meeting attendance system for Telkom Indonesia using GhostFaceNet, achieving 98% accuracy during validation and testing phases across a 14-person engineering team's production deployment.

  GhostFaceNet Face Recognition Computer Vision Python

• 04

  Avatar Generation for Video Learning

  Automated lip-sync and realistic avatar movement pipeline for e-learning videos using Wav2Lip and Sf3d, cutting manual video editing time by 50% in production at Telkom Indonesia.

  Wav2Lip Sf3d Video Generation Deep Learning

• 05

  Bank Deposit Propensity Model

  Analyzed feature correlation in bank client data and built a propensity model using Gradient Boosting and Logistic Regression, achieving 93% accuracy in predicting deposit product applications. Also processed 3M+ rows of cybersecurity attack data for anomaly detection.

  Gradient Boosting Logistic Regression Tabular Data Scikit-learn

• 06

  EEG Grip Strength Prediction

  Preprocessed EEG signals and identified channel correlation to predict grip strength and user context (VR vs. real-world). Random Forest and XGBoost ensemble achieved 96% accuracy.

  XGBoost Random Forest EEG Signal Processing

• 07

  Denoising Autoencoder

  Image denoising model built with PyTorch, trained to remove noise from corrupted images. Achieved MSE loss of 0.0019 on the test set.

  PyTorch Autoencoder Computer Vision Deep Learning