Thai Vu Nguyen

Technical Skill

  • Programming Language: Python, C/C++, C#.
  • Framework: Tensorflow, Pytorch, CUDA.
  • ML Engineering: Google Cloud, Docker, ML Flow, Digital Ocean.
  • Software Engineer: Flask, Streamlit, HTML/CSS.
  • Tools & Platforms: Git/Github, SSH.

Project

Quantization Neural Machine Translation [Link]

  • Developed English-Vietnamese translation model using Transformer architecture from "Attention is All You Need.".
  • Trained on the PhoMT dataset with 3 million sentence pairs, achieving a BLEU score of 0.26 on the test set.
  • Applied asymmetric post-training quantization to multi-head attention layers, converting float32 parameters to int8, resulting in a 20MB reduction in model size.

GPU Computation [Link]

  • Accelerated Convolutional Neural Network inference (2D convolution, pooling, fully connected layers) using CUDA C++ for fast prediction.
  • Implemented parallel matrix operations, including LU decomposition and image processing (edge detection, sharpening, and blurring) on GPU.
  • Applied Seam Carving for intelligent resizing and Histogram Equalization for image enhancement.

Emoji Search Engine [Link]

  • Developed a search engine that allows users to find emojis based on input text.
  • Improved data quality by processing and augmenting an emoji dataset. Utilized GloVe embeddings with XGBoost for emotion prediction.
  • Deployed web application on DigitalOcean using Flask backend.

Hockey Match Analysis [Link]

  • Developed a real-time hockey match analysis web app with Flask, Streamlit, and Docker for deployment.
  • Used NHL APIs to retrieve and analyze hockey data with statistical, cleaning data, and visualization techniques.
  • Trained an XGBoost model to predict shot goal probability and used Comet for data version control.

Generative Adversarial Network [Link]

  • Applied image generative techniques, including GAN, Semi-GAN, Conditional GAN, Progressive Growing GAN, and Cycle GAN, to generate high-quality synthetic images.
  • Key technologies: Python, Tensorflow/Keras.