Tai Nguyen Phu

What I have done:

• Fine-tuned a Sentence-BERT model (bert-base-multilingual-cased) on a 100K+ Vietnamese legal documents using CachedMultipleNegativesRankingLoss for domain-specific semantic retrieval.
• Achieved significant performance boost on MTEB evaluation benchmark (NDCG@10: 60.4%, MAP@10: 53.6%) compared to pre-trained baseline (NDCG@10: 0.023%, MAP@10: 0.014%).
• Designed and deployed a scalable semantic search pipeline using FAISS for real-time legal document retrieval.
• Built a user-friendly web interface with Gradio to facilitate interactive query testing.
• Containerized the full system with Docker to ensure reproducibility, portability, and seamless deployment.

What I have done:

• Developed an AI-powered book recommendation system that leverages semantic search and zero-shot classification to suggest books based on user input.
• Utilized pre-trained transformer models to generate embeddings for book descriptions, enabling semantic similarity comparisons.
• Implemented zero-shot classification to categorize books into genres without labeled training data.
• Built an interactive web application using Gradio, allowing users to input descriptions and receive personalized book recommendations.
• Conducted exploratory data analysis and sentiment analysis to enhance recommendation relevance.

What I have done:

• Developed a Gen AI assistant to streamline job applications by automating 80% of the cold email generation process and providing a tailored chatbot for job application guidance based on job postings and applicant resumes.
• Implemented FAISS-based semantic search for efficient retrieval of job descriptions and resumes.
• Integrated LangChain to enhance AI-driven responses with contextual understanding.
• Designed and deployed a Streamlit UI, ensuring seamless user interaction.

What I have done:

• Built a dense passage retrieval system using dual-encoder transformers (Hugging Face), enabling semantic search over large QA corpora.
• Achieved ~1% improvement in top-k retrieval accuracy over the original DPR baseline.
• Leveraged FAISS for fast vector similarity search, ensuring scalable, real-time response even under memory constraints.
• Optimized training and inference on 16GB GPU, using batch scheduling, mixed precision, and memory-efficient data loaders.

What I have done:

• Developed a Scene Text Recognition pipeline using YOLOv11m for text detection and CRNN (ResNet34) for recognition, achieving ~88% precision on the ICDAR2003 dataset.
• Developed a scalable OCR API with FastAPI and Ray Serve, enabling autoscaling, GPU acceleration, and real-time text extraction.
• Created an interactive Streamlit web application, allowing image processing via file upload or URLs for instant text recognition.
• Implemented transformer-based text processing and CTC loss for sequence modeling, ensuring accurate recognition of text in natural scene images.

What I have done:

• Developed a real-time neural style transfer pipeline using a lightweight feed-forward CNN trained on MS-COCO and stylized with artworks (e.g., Monet, Van Gogh).
• Integrated OpenCV for live webcam streaming and real-time frame capture, enabling dynamic style application on video feeds.
• Deployed the system via Streamlit to provide an intuitive web interface with instant style switching and preview rendering.
• Implemented frame-by-frame normalization, resizing, and transformation pipelines to maintain visual fidelity across different devices and resolutions.

What I have done:

• ...

What I have done:

• Developed a machine learning-powered salary prediction model for Data Science jobs, trained on real-world Glassdoor job market data.
• Engineered and optimized features using EDA, one-hot encoding, label encoding, and feature importance analysis.
• Trained and compared multiple ML models (XGBoost, Random Forest, Decision Tree, Gradient Boosting, Linear Regression, SVM), achieving R² = 0.815 with XGBoost.
• Implemented hyperparameter tuning with GridSearchCV & RandomizedSearchCV to maximize prediction accuracy.
• Built and deployed an interactive Streamlit web application, allowing users to input job details and receive salary predictions.