Nguyen Cong Phat (James)

ViHateT5 Reimplementation – Vietnamese Hate Speech Detection

Reimplementation & Improvement of ViHateT5 for Vietnamese Hate Speech Detection | Group Project

Role: Leader – Course: DS200.Q21 – Big Data Analysis (UIT – VNU-HCM).
Led the full reimplementation of the ACL 2024 ViHateT5 paper, designing and coordinating a unified text-to-text Transformer pipeline for multi-task Vietnamese hate speech detection across ViHSD, ViCTSD, and ViHOS datasets.
Developed a large-scale auto-labeling pipeline using ViSoBERT to annotate 10M+ VOZ forum samples with 97.5% agreement, enabling continual domain-specific pretraining of ViT5 using span-corruption objectives.
Implemented and benchmarked 8 BERT-based encoders and 3 T5-based Seq2Seq models, reproducing and improving paper-level performance while deploying both FastAPI and Streamlit inference applications.

Python Transformers PyTorch NLP T5 BERT HuggingFace FastAPI Streamlit

Code Repository Models & Datasets Original Paper

Zero-shot Object Counting with Good Exemplars (VA-Count Extension) | Group Project

Course Project at UIT – VNU-HCM .
Spearheaded the implementation of an enhanced zero-shot counting framework, integrating Gemini-based Rich Prompts and CLIP re-ranking to improve exemplar quality, achieving a MAE of 17.80 on the FSC147 dataset.
Integrated and Benchmarked YOLO-World as a high-speed open-vocabulary detector, successfully reducing inference latency by 2.5× (down to ~0.6s/image) while maintaining competitive counting accuracy.
Optimized and Deployed a robust end-to-end pipeline via Streamlit, featuring real-time density map visualization and automated positive/negative exemplar extraction for generalized counting in unseen object classes.
References: Based on the VA-Count (ECCV 2024) methodology with extensions from Rich Prompts (Zhu et al., 2025).

Python PyTorch YOLO-World GroundingDINO Large Language Models Computer Vision Streamlit

Code Repository

Big Data Analysis Lab Workspace – Hadoop MapReduce & Apache Pig Pipelines

Role: Individual Project – Course: DS200.Q21.1 – Big Data Analysis Lab (UIT – VNU-HCM).
Designed and implemented distributed data processing pipelines using Java MapReduce to analyze structured movie-rating datasets, including aggregation, filtering, and transformation workflows.
Built Apache Pig Latin scripts for large-scale hotel-review text analytics, performing preprocessing, stopword filtering, and frequency-based keyword extraction using Hadoop-style workflows.
Organized a modular multi-lab big data workspace supporting Java MapReduce, Hadoop Streaming (Python), Pig scripts, and optional pandas-based validation pipelines for reproducible experiments.

Java Hadoop MapReduce Apache Pig Python Pandas Big Data

Code Repository

Wildfire Semantic Segmentation and Alert System for UAVs via Deep Learning | Group Project

Course: CS406.Q11 – Image Processing and Applications (UIT – VNU-HCM).
Architected and optimized a wildfire detection framework using U-Net architectures, achieving a state-of-the-art IoU of 0.7139 on the FLAME dataset for high-precision boundary delineation.
Implemented and validated a specialized training pipeline featuring advanced data augmentation and a 3-loss function strategy to prioritize recall, resulting in an F2-score of 0.8162 for reliable fire alerting.
Engineered and deployed a spatio-temporal tracking logic within a real-time monitoring system, maintaining geographic localization errors under 50 meters for effective UAV-based emergency response.
Special thanks to the team at Sakocpo/Wildfire-Detection-Deep-Learning for their foundational codebase and research contributions.

Python PyTorch Computer Vision Semantic Segmentation U-Net Albumentations Gradio

Code Repository

Face Mask Classification – Traditional ML vs CNN

Face Mask Classification – A Comparative Study of Traditional ML and CNNs | Group Project

Role: Leader – Course: CS231.Q11 – Introduction to Computer Vision (UIT – VNU-HCM).
Led and coordinated the end-to-end development of a face mask classification system, defining the experimental methodology and driving a comparative evaluation between traditional ML pipelines (HOG/LBP + SVM, KNN, Random Forest) and CNN-based models.
Designed, implemented, and deployed the full computer vision pipeline, including data preprocessing, feature extraction, hyperparameter optimization (Optuna, Keras Tuner), and real-time inference via OpenCV (YuNet) and Flask, achieving up to 98.99% test accuracy and >25 FPS real-time performance.

Python Computer Vision OpenCV Machine Learning CNN HOG LBP Flask

Code Repository Results Demo

Hate Speech Detection and Highlighting for Vietnamese (HARE)

Rationale-Augmented LLM Training for Targeted Hate Speech Detection in Vietnamese Social Media | Group Project

Role: Leader – Course: IE403.Q11 – Social Media Mining (UIT – VNU-HCM).
Initiated and led the HARE research framework by formulating the core hypothesis on rationale-based explainable hate speech detection and designing the fine-tuning strategy for Qwen2.5-3B using QLoRA (4-bit) and Chain-of-Thought prompting.
Constructed and curated an enhanced ViTHSD dataset by enriching 2,333 samples with rationale and implied-statement annotations, and implemented a two-stage semantic alignment pipeline (HARE) combining label supervision with reasoning-based learning.
Evaluated and deployed the system through a full-stack application (FastAPI + React/Vite), achieving F1 = 0.6026 on the ViTHSD test set and delivering a +17% F1 improvement on political hate speech compared to vanilla Qwen models.

Python NLP Data Science LLM Explainable AI Qwen2.5 QLoRA FastAPI React

Code Repository Results Demo

Vietnamese Named Entity Recognition (VLSP 2016)

Vietnamese Named Entity Recognition (NER) | Group Project

Role: Leader – Course: CS221.Q12 – Natural Language Processing (UIT – VNU-HCM).
Led and designed the end-to-end Vietnamese NER project by defining the experimental framework and implementing three sequence labeling models (HMM, CRF, BiLSTM-CRF) on the VLSP 2016 dataset (CoNLL format).
Evaluated and benchmarked model performance using token- and entity-level metrics, with the CRF model achieving Acc = 0.9904, Token F1 = 0.9901, Non-O F1 = 0.9076, and Span F1 = 0.9191 on the test set.
Developed and deployed a Flask-based interactive demo for model comparison, real-time entity highlighting, and result visualization, ensuring reproducibility through archived models and training artifacts.

Python NLP HMM CRF BiLSTM-CRF PyTorch Flask

Code Repository Results Flask Demo

VPBank Customer Service Usage Analysis | Personal Project

Analyzed over 113K+ customer transactions and built interactive Power BI dashboards to visualize service usage patterns and revenue flows.
Segmented customers by geography, asset value, and service adoption to identify “Gold” customers contributing 78% of total assets.
Implemented customer profiling to classify Regular, Silver, and Gold segments, uncovering high cross-sell potential.
Generated actionable insights and proposed data-driven campaigns that supported a +20% quarterly revenue growth target.

SQL Power BI Data Analytics

View Project Files

FPT Play Sentiment Analysis | Personal Project

Developed an end-to-end pipeline to extract feedback data from MongoDB and transform it for MySQL storage using Python.
Implemented a Vietnamese text classification model using phoBERT, achieving a 92% F1-score on validation data.
Created an interactive dashboard to visualize customer sentiment and inform service quality improvement initiatives.

Python MongoDB MySQL phoBERT

View Project Files

Loan Approval Prediction | Group Project

Role: Feature Engineering \& Model Development Lead.
Led EDA and engineered new features through correlation-based elimination and data cleaning to enhance signal quality.
Developed and optimized tree-based models (XGBoost, LightGBM, CatBoost) via 5-fold cross-validation, achieving an F1-score of 0.8939 and AUC of 0.9631.
Authored and presented the final report, translating technical findings into clear narratives for academic evaluation.

Python XGBoost LightGBM CatBoost

Code Repository

Realtime Stock Analysis Tool for Vietnam Market 🇻🇳 | Personal Project

Developed a real-time Streamlit dashboard analyzing capital flow and technical indicators in the Vietnamese stock market using vnstock API.
Visualized inflow/outflow trends, bid-ask ratios, and capital movement heatmaps through matplotlib and seaborn.
Optimized user interactivity with dynamic ticker input, time range selection, and statistical summaries for quick insights.

Python Streamlit vnstock Data Visualization

Code Repository

My Gallery

A minimalist online gallery to display favorite photos.
Built for fun, practice, and to share personal photography.
Responsive layout with smooth animations.

HTML CSS JavaScript

Code Repository Live Demo

Bosch Vietnam Demo Website

Personal learning project inspired by Bosch’s values.
Multi-page layout (Home, About, Products, Blog, Contact, etc).
Responsive design using Bootstrap 5, WOW.js, and Owl Carousel.

Bootstrap 5 WOW.js Owl Carousel Web Design

Live Demo

Python Rock–Scissors–Paper Game (OpenCV) | Personal Project

Developed an interactive Rock–Scissors–Paper game using Python + OpenCV for real-time hand gesture recognition via webcam feed.
Implemented an automated opponent system and visualized player–computer outcomes using OpenCV overlays.
Designed modular code with a custom hand detection library for scalability and future GUI enhancements.