• Developed an expression transfer model and pipeline enabling dynamic, controllable emotion transitions (joy, sadness, anger) in video generation for human characters.
• Designed and deployed a real-time audio-driven talking head system (lip-sync, singing) preserving identity, head pose, and facial expression from a single audio stream.
• Built a real-time Stable Diffusion video generation pipeline producing AI avatar speech video from audio input.
• Developed a real-time child attention monitoring system using facial-landmark head pose estimation and screen engagement tracking, plus a skeleton-based gym exercise recognition system.
• Architected a multitask VLM pipeline for complex visual understanding, integrating optimised inference and VLM fine-tuning for spatial–visual reasoning.

• Module owner for Face Analysis and Perception: face detection, face recognition, object detection, segmentation, text-to-face synthesis, text-to-video.
• Authored and published research at NeurIPS 2022; filed 3 patents covering gaze prediction, face recognition training, and edge DMS.
• Developed a Neural Architecture Search pipeline to auto-generate optimal object detection architectures for CPU-constrained MHU devices.
• Performed post-training quantization and QAT (FP32 → INT8), achieving significant model compression with minimal accuracy loss across face and detection models.
• Deployed CV models across 10+ hardware targets — NVIDIA Xavier, Qualcomm QCS610/SA8255P, KingFisher, VinSmart mobile — using TensorRT, SNPE, QNN, CVFlow, NCNN, MNN, and TFLite.
• Built the production Driver and Occupancy Monitoring System, recognised Top 5 at AI4VN 2022 and a CES 2024 Innovation Award Honoree.
• Delivered lightweight face analysis networks (FaceID, Masked FaceID, camera effects) for VinSmart mobile devices.

• Taught Python programming and robotics (Mindstorms) to high school students, including preparation tracks for US university admissions.

CGPA: 3.21/4.0

Thesis: QC-StyleGAN2: Controllable Output Quality Generative Adversarial Networks — graded 9.79/10 (Excellent). Supervised by Professor Cuong Pham.