Glad to have an AI/Data Project together!

I am part of a six person team developing what we call as ’Zupa App’. We brings public safety especially women and children safety into an application, as a companion app that can deliver safety tips, SOS button, and make enforcer’s task easier. As Machine Learning head developer, we made in app CNN model that could detect four types of violence using audio input activated with SOS button with 80% + accuracy.

Banany is a portable, machine learning powered backup device for photographer, also powered with wireless network, access point, and NAS capability. Banany uses pretrained model - weighted using MSCOCO - to classify most dominant object in backup photos. This project awarded with gold medal plus special award in Internet of Things from Japan Design Invention Expo 2019, and bronze medal plus special award from Malaysia Technology Expo 2018.

Using Java and Python, me and my 2 colleagues created a system for Indonesian dine-in home dish restaurant (’Warteg’) so consumer have more information in lacking of transparency by the cashier. It took an image of our plate containing rice and condiments, then using mask RCNN to identify condiments and sum up the price taken from database.

Provide and review basic mathematics, machine learning, deep learning, and AI deployment materials for Digital Talent Scholarship. Thankfully with my deeplearning.ai specialisation I am able to design ideal materials for AI starters but with industrial needs.

Sirkadian is a startup I co-founded with medical students of Udayana University with goal to deliver healthy lifestyle as a trend with four core, mental, physical, diet, and healthy sleep. In this role I am able to process food and ingredients database, also verify the authenticity and validity of AI algorithms used in food necessity and food recommendations system.

In Intelligent System and Machine Learning Laboratory, We create electrical bill prediction system, my role is to deploy the machine learning model as a website and clean the initial data. We also have a plan to publish this as an academic paper.

EfficientDet is a relatively new object detection network, hence all of its poteniality has not been explored. EfficientDet as an object detection network works well, but what if the network faced with items that similar in shape yet different class? Like a box of stawberry milk and a box of chocolate milk? This challenges smart-retail implementation in daily life. Me and my colleagues tested popular object detection models in several Indonesian products available in store.

In this project I helped my senior's thesis to detect faces and automate online class attendance. Face detection model was using RCNN with two classes (attend and absent), and attendance automated using Python scripts. This thesis graded A (excellent) after defence.

I helped my senior's thesis to detect non-porn and porn content in twitter's tweets(1). I helped to fine-tune, and analyze the traning result. I also helped my other senior to prepare the data, create, and evaluate Bidirectional LSTM model to detect IndiHome (Indonesian ISP) customer feedback's sentiment collected in form of twitter's tweets(2). As a result, both of the thesis got excellent result during thesis defence.

I currently learning MATLAB and GNU Octave to enhance my digital signal processing skill and expand my code experience outside of general purposed languages, like C and Python to more scientific purposed languages like MATLAB. I think that would be good to share my learning journey and get code samples along since it is rare around internet. Usually I update this repository every one or two weeks in the weekend.

In this Laboratory, we believe in continous learning and we support each other. Hence weekly we have a crash course (around 2 hours webinar per week) talking about technology stack and showcasing its implementation. I mainly share mathematics and machine learning materials such as predict Titanic's passenger survivability using linear regression and data analysis, and create binary classifier from spectogram using K-Nearest Neighbour algorithm.

Using the power of machine learning, this repo tends to recover colour from greyscale image. Usually I will try neural network, but as i plot it, it shows towards a regression problem, polynomial to be precise. In my experience, the NN does not work pretty well with linear problem like this but since we need to figure out a lot of data, that worth a try. But if the result is not getting better (currently stuck in 12% acc), i will consider changeing to polynomial regression.