An AI or data project together?
I build, deploy and teach machine learning, and I help research teams with code, analysis and manuscripts. Tell me what you are trying to do.
How I got into AI
My journey with machine learning started while I was preparing for the Japan Design Invention Expo. I was building an embedded system for photographers to store their photos. The device was already at version 2 and I was out of ideas for improving it beyond transfer speed and network capability. The answer was intelligence. I implemented TensorFlow, starting from some code on GitHub, so the device could recognise photos and sort them into folders by what it saw inside. I fell in love with machine learning after that, and started digging into data mining, Python, statistics and courses.
What I did for Japan motivated me to go deeper into AI, machine learning and deep learning. I want to turn my imagination into accepted science, or into something real. I like to study, adapt and treat a challenge as a quiz that pays off at the end.
Tech stack
Levels are my own judgment of my experience and confidence, except where I say licensed.
Strong
- Deep learning with TensorFlow (licensed)
- Python (licensed)
- Mathematics for machine learning (licensed)
- Deep learning deployment
- AI education
Working knowledge
- AI and machine learning theory
- Data analysis and preprocessing
- MATLAB
Still learning
- C and C++ for AI
- R
Set the data, set the main goal, evaluate, deploy.
Past projects
Mostly from my undergraduate years and the first years after. My current research is on the main page.
-
Violence audio detection with a convolutional neural network
I was one of six people building Zupa, a companion app for public safety, especially for women and children, with safety tips and an SOS button. As machine learning lead, I built the in-app CNN that detects four types of violence from audio once the SOS button is pressed, at over 80% accuracy.
-
Photo classification in BANANY, Backup Anywhere Anytime
BANANY is a portable backup device for photographers with wireless networking, an access point and NAS capability. It uses a pretrained model, weighted on MSCOCO, to classify the dominant object in each backed-up photo. The project won a gold medal and a special Internet of Things award at the Japan Design Invention Expo 2019, and a bronze medal and special award at the Malaysia Technology Expo 2018.
-
Automatic billing for a warteg using Mask R-CNN
With two classmates, I built a system for Indonesian home-style restaurants (warteg) so customers can see what they are charged. It photographs a plate of rice and side dishes, uses Mask R-CNN to identify each item, and sums the prices from a database. Java and Python.
-
Deep learning deployment course material
I wrote and reviewed material on basic mathematics, machine learning, deep learning and deployment for the Digital Talent Scholarship, designed for AI beginners with industry needs in mind. It drew on my deeplearning.ai specialisation.
-
AI algorithm review
Sirkadian is a startup I co-founded with medical students of Udayana University, built around healthy lifestyle: mental, physical, diet and sleep. I processed the food and ingredient database and checked that the AI algorithms behind food recommendations were valid.
-
Electrical bill prediction with time series
The lab built an electrical bill prediction system. I cleaned the initial data and deployed the model as a website. We planned to publish it as a paper.
-
Retail product classification with EfficientDet
EfficientDet detects objects well, but what happens with items of the same shape and different class, like a box of strawberry milk and a box of chocolate milk? That gap holds back smart retail. My colleagues and I tested popular detection models on Indonesian store products.
-
Online class attendance with R-CNN
I helped build face detection with two classes (attend and absent) and Python scripts that automate online class attendance. The thesis was graded A after the defence.
-
Tweet content detection and customer-feedback sentiment with BiLSTM
For the first, I helped fine-tune a model that separates adult from non-adult tweets and analysed the training results. For the second, I prepared the data and built and evaluated a bidirectional LSTM that classifies the sentiment of IndiHome customer tweets. Both theses had excellent defences.
-
MATLAB and GNU Octave learning repository
I learned MATLAB and GNU Octave to strengthen my digital signal processing and to move beyond general-purpose languages like C and Python. I shared the code samples publicly, because they are rare online.
-
Weekend crash course at Laboratorium Dasar Komputer
The lab ran a weekly two-hour webinar on a technology stack and how it is used. I mostly taught mathematics and machine learning: predicting Titanic survival with linear regression and data analysis, and building a binary classifier from spectrograms with k-nearest neighbours.
-
Recovering one missing colour channel without deep learning
This repository tries to recover a missing colour channel of an RGB image with classical machine learning only. When I plotted the data it looked like a polynomial regression problem, where neural networks tend to do poorly. A first attempt stalled at 12% accuracy, so polynomial regression was the next step.