Software Engineer at Visa

github

Huy Lai  • Peerlist

https://bvcog-d102bb1214eb.herokuapp.com/users/sign_in

BVCOG Contract Management System

<p>This project focuses on improving the training of CLIP (Contrastive Language-Image Pretraining) models by optimizing the global contrastive loss for bimodal contrastive self-supervised learning. Self-supervised learning (SSL) has gained prominence for its ability to generalize across downstream tasks in areas such as natural language processing and computer vision. Among SSL frameworks, Contrastive Learning (CL) has proven effective by maximizing the similarity between positive pairs and minimizing it between negative pairs. While CLIP has demonstrated success in aligning image and text representations, challenges persist, such as slow convergence in large-scale bimodal datasets. Participants are tasked with accelerating global contrastive loss optimization and enhancing model performance on provided benchmarks. Trainig of CLIP models used a 100k subset of the Conceptual Captions 3M dataset for training and validate the models on MSCOCO and ImageNet datasets. They must evaluate model performance based on retrieval accuracy and zero-shot classification metrics. Models are restricted to using ResNet-50 and DistilBERT as encoders, with fixed hyperparameters, and must compare at least two optimizers and three loss functions. Deliverables include model code, trained models, and a detailed report covering experimental results, all adhering to specified guidelines. Evaluation criteria include experimental breadth, report quality, presentation, and innovative ideas.</p>

Improving CLIP Training

Collaborated with a team to develop a predictive model for in-hospital mortality using eICU data. Trained an XGBoostClassifier model acheiving an Area Under the Receiver Operating Characteristic curve (AUC-ROC) Score of 0.89288. Conducted rigorous data preprocessing, feature selection, and hyperparameter tuning. Analyzed feature importance for mortality classification.

Machine Learning for Mortality Prognosis

A Node.js and React project that creates a Point-of-Sale webpage for a Restaurant located at the Memorial Student Center at Texas A&M University. Our specific restaurant was Chick-fil-A. Using Google OAuth to authenticate administrator logins, adequate role-based access control is enforced. This project also uses a Google Translate API to ensure proper internationalization. Finally, this project uses a weather API to display weather information as well as recommend certain items during cooler weather. A custom REST API to communicate between the front-end and the back-end of the application.

Point of Sale System

<p>Retrieving specific information from vast video data on platforms like YouTube presents a formidable challenge due to the limitations of traditional metadata-based methods. These methods often struggle to capture the intricacies of video content accurately. Additionally, the sheer amount of content uploaded to YouTube ensures that a comprehensively nuanced search would take an unreasonable amount of time and effort. By storing YouTube transcripts and developing a dedicated website, this solution aims to enhance search experiences, optimizing performance while maintaining a balance between speed and accuracy. The system enables flexible searches with sorting and filtering capabilities, empowering users to efficiently locate videos based on specific words or phrases they recall. Evaluation encompasses rigorous unit testing, integration testing, user acceptance testing, and user studies to ensure both functionality and user satisfaction. Overall, the website provides an effective tool for navigating and accessing video content with precision and ease.</p>

ScriptSearch 🔎

Implemented a parallel version of Strassen's matrix multiplication algorithm using OpenMP directives in C/C++. The project aimed to leverage parallelism to improve the efficiency of large-scale matrix multiplication tasks. Developed a scalable parallel algorithm based on Strassen's matrix multiplication technique, utilizing OpenMP directives for parallelization. Optimized the code to efficiently distribute workload across multiple threads, achieving significant speedup compared to the sequential version. Conducted rigorous testing and performance analysis to evaluate the scalability and efficiency of the parallel implementation across various input sizes and hardware configurations. Documented the design, implementation details, and performance evaluation results comprehensively for future reference and potential publication.