Projects

Things I built at work and on my own, in AI, machine learning, and computer vision.

• MARSHALL

  Telkom Indonesia

  A platform that helps Telkom Indonesia create and review learning modules with LLMs. 1,406 employees used it. I built the RAG pipeline, which raised answer factuality from 20% to 90%. With a team of 14, we cut manual work by 40% in six months.

  GPT-4o Llama 3.2 RAG FastAPI

• Live2D AI companion

  Personal project

  A chat companion with a Live2D avatar. You talk to it by voice or text, and an LLM writes the replies. I built it to learn frontend development.

  Live2D LLM Voice chat Frontend

• Face recognition attendance

  Telkom Indonesia

  An attendance system for online meetings. It uses GhostFaceNet and reached 98% accuracy in validation and testing.

  GhostFaceNet Face recognition Python

• Avatar generation for video learning

  Telkom Indonesia

  A pipeline that makes talking avatars for e-learning videos. Wav2Lip handles lip sync and Sf3d builds the avatar. It cut manual video editing time by 50%.

  Wav2Lip Sf3d Video generation

• Defect product detection

  Research, Yogyakarta State University

  A YOLOv8 model that finds defects in manufactured products from a live camera. I built it as learning material for a robot vision course. It shows Grad-CAM maps of what the model looks at, and it exports to NCNN for edge devices.

  YOLOv8 OpenCV Grad-CAM NCNN
  View code on GitHub

• NeetCode GPT

  Personal project

  A GPT built from scratch in NumPy, one part at a time, from my NeetCode ML course work. So far it covers neurons, backpropagation, gradient descent, MLPs, and softmax.

  NumPy Neural networks From scratch
  View code on GitHub

• Autonomous drone with computer vision

  Dreamagination

  A competition drone that finds objects, picks them up, and avoids obstacles. YOLOv3s runs with TensorRT on a Jetson Nano, and odometry runs on ROS. It picked objects correctly 90% of the time in real tests. Our team placed 2nd at FIRA Air SimulCup 2022 and KRTI Regional 2023.

  YOLOv3 TensorRT ROS Jetson Nano

• Bank deposit propensity model

  Freelance

  A model that predicts which bank clients will apply for a deposit product. Gradient Boosting and Logistic Regression reached 93% accuracy.

  Gradient Boosting Logistic Regression Scikit-learn

• EEG grip strength prediction

  Freelance

  A model that reads EEG signals to predict grip strength, and whether the user is in VR or the real world. A Random Forest and XGBoost ensemble reached 96% accuracy.

  XGBoost Random Forest EEG

• Denoising autoencoder

  Personal project

  A PyTorch autoencoder that removes noise from images. It reached an MSE loss of 0.0019 on the test set.

  PyTorch Autoencoder Computer vision