List of Past Projects

Project 6: Recommend services using Large Language Models

To improve customer experience, a prototype chatbot was constructed to perform event-based recommendation tasks and is in the process of integrating with the existing business operation. This chatbot used OpenAPI for processing natural language.

Keywords: OpenAPI, Content-based filtering, Apriori, Clustering

Project 5: Predict different respiratory diseases using multi-modal data and deep learning models

To improve the predictive power, an algorithm that can integrate different modalities of data, such as text and speech, transformers were utilized to extract audio features along with three different machine learning classifiers in order to conduct a prediction task for respiratory disease.

Keywords: Wav2Vec2.0, XGBOOST, Light GBM, CNN models

Project 4: Predict admitted emergency patient outcomes

The innovation in technologies related to health facilities today is increasingly helping to manage patients with different diseases. One of the most challenging issues is to accurately predict the survival outcome for mechanically ventilated patients in intensive care units (ICUs). The real challenge is to diagnose patients with more diagnostic accuracy and in a timely manner, followed by prescribing appropriate treatments and keeping prescription errors to a minimum. This project constructed machine learning models to predict patients’ outcomes and deployed a cloud-based ICU prediction system empowered with a set of machine learning models, including Bagging (BGG), eXtreme Gradient Boosting (XGB) based system model gives 90+% accuracy.

Technical publication available to access: https://doi.org/10.3389/fmed.2024.1398565

Product: Web Application

Keywords: AWS EC2, Postgre SQL, Airflow, Bagging, XGBOOST, eXtreme Gradient Boosting, Decision Tree

Project 3: Classify household waste

An application was built to classify 4 different types of waste, namely organic, plastic, landfill, and refundable. Different teams were working from collecting data to train computer vision models to perform the classification tasks, from optimize machine learning models to deploy model on AWS cloud-based.

Product: Web Application

Keywords: AWS EC2, Postgre SQL, Airflow, Streamlit, Ultralytics Yolo

Project 2: Evaluate fine-tune large language models on Q&A systems

T5, BART, BERT, RoBERTa models were selected to evaluate its performance of factuallity questions. Different teams construct questions and ground truth/label/answers, trained T5 model, and evaluate its performance. F1 and BLEU scores were utilized for evaluating the performance of those trained models. The output of this work can be used to build a Q&A application for educational institutions.

Keywords: T5, BART, BERT, RoBERTa, F1, BLEU

Project 1: Evaluate hyperparameter optimization methods

Hyperparameter methods, including Random Search, Bayesian Optimization, and Particle Swarm Optimization were examined their impacts on the classification accuracy of machine learning models, namely XGBOOST, Random Forest, Support Vector Machine, and K-nearest neighbors. This gives experience to machine learning operators in ways to optimize ML models while utilizing less required resources.

Keywords: XGBOOST, Random Forest, Support Vector Machine, and K-nearest neighbors, Random Search, Bayesian Optimization, and Particle Swarm Optimization