Projects

Developing a motion scoring system to help user in performing physiotherapy exercises. The system utilizes a production model of Google's PoseNet for pose estimation and engineered for scoring similarity and counting movement repetitions.

Developing a hybrid (content-based and collaborative) reccomendation system to help match freshmen college students entering Fasilkom UI. The recommendation system utilizes a modified LightFM deployment to account for the reciprocity of the system.

Developed a new pipeline for age estimation & gender classification from photos extracted from social media. As required, the pipeline does not use pretrained model and involves face detection alignment and an estimator trained using multi-task learning.

Conducted an experiment to study the effectiveness of replacing the usual convolutional stem of ResNet with a patch-based stem. As demonstrated by ConvNext, the experiment found that the patch-based stem is slightly better.

Developed a multimodal and semisupervised model for identifying useful/informative tweets during a flood disaster. The model was trained on both image and textual data and was further improved upon by an unlabled set of data using semisupervised learning.