Sastra Bali.

Learning through the camera

Sastra Bali, also called Sabi, is a Bangkit capstone app for learning the Balinese language and script. Alongside learning materials and quizzes, its scanning feature helps learners identify Balinese aksara from an image.

I worked on the machine learning side as part of a three-person ML team, collaborating with the Android and cloud teams to turn a character classifier into an app feature.

Building the dataset and model

We combined a public Balinese aksara dataset with handwritten samples we created ourselves. The training workflow brings those sources together, augments the images, and trains a convolutional neural network in TensorFlow. Separate notebooks cover augmentation, model training and evaluation, and metadata preparation.

The work included iterating on the model and checking its classification results before preparing it for the mobile team. Our original real-time video OCR ambition narrowed into a scan-and-classify flow that we could deliver within the project.

Making the model usable on Android

The handoff included a TensorFlow Lite model and metadata to support integration into the Kotlin Android app. That made the model’s input and output contract part of the implementation, alongside the network itself.

What I learned

This project connected dataset preparation, model iteration, and mobile deployment. The most useful lesson was that a trained model becomes a product feature only when another team can integrate it: preprocessing, labels, model format, and metadata all matter.

The machine learning repository contains the data preparation and model workflow; the organization also holds the mobile and backend components.