- Spearheaded the design of performance metrics and optimization strategies for News and Music recommendation systems, leveraging neural collaborative filtering (CF) and content-based (CB) feature extraction. Automated end-to-end pipelines, improving recommendation accuracy and reducing manual intervention. - Developed Large Language Model (LLM)-based Retrieval-Augmented Generation (RAG) pipeline, enhancing semantic search capabilities for Chatbot and Question Answering systems, boosting response relevance and user satisfaction. - Designed and implemented a highly accurate Vietnamese spell-correction model by leveraging word-level and character-level BERT architectures, significantly improving the text quality in downstream NLP applications. - Developed and optimized Named Entity Recognition (NER), tagging, and classification models tailored for Vietnamese news content, enhancing contextual understanding and achieving improved precision in text analytics for media-related tasks. - Developed an automated video processing service, including precise intro/outro trimming and intelligent thumbnail selection, streamlining content management and enhancing user experience for video platforms.

Collaborate closely with customers in England to analyze requirements. Design the system to detect droplet images. Classify and measure the size of the droplets.

Conducted advanced research and development on core Computer Vision solutions for Intelligent Video Analytics, focusing on Intrusion Detection, Face Recognition, Anti-Spoofing, License Plate Recognition, and OCR, driving innovation across key surveillance systems. Optimized serving with Deepstream, TensorRT/ONNX, and Qualcomm’s Hardware. Collaborated with Technical Leadership to enhance Face Recognition models for NIST vendor testing, implementing state-of-the-art neural network backbones and data augmentation techniques to improve accuracy, especially for side-view angles, on large-scale datasets.

Engineered an optimized Gradient Response Maps algorithm for shape-based matching, incorporating a novel hierarchical search mechanism to increase matching speed and accuracy for high-throughput inspection systems. Optimized the PADIM algorithm for anomaly detection by customizing its architecture and leveraging einsum for Mahalanobis distance calculation. Integrated supervised and unsupervised algorithms into an innovative framework for anomaly detection.

Customized OCR systems to meet diverse client requirements. Collaborated directly with customers to analyze needs. Designed tailored solutions for optimized document recognition workflows.