Generative AI
Personality-Driven AI Avatar Bot
AI avatars with distinct personalities that join a group chat — generated end to end, from character identity through to speaking video.
Dec 2023 – Jan 2025
☕️Dat Viet Thanh Nguyen
Senior AI Research Engineer
H2 Corporation
Making vision and generative models small enough to ship.
About Me
I am a Senior AI Research Engineer at H2 Corporation, working on multitask vision-language pipelines for complex visual understanding. For the past 6 years I have moved between research and production: publishing at NeurIPS and reviewing for top-tier vision conferences on one side, and shipping computer vision models onto edge SoCs, phones, and cloud services on the other.
My work has spanned face analysis and perception, object detection, generative models, and real-time video AI — from lightweight FaceID networks running on VinSmart phones, to a driver monitoring system recognised at AI4VN and CES, to real-time audio-driven talking heads and diffusion-based avatar video. A recurring thread is making models small and fast enough to be useful: neural architecture search for CPU-constrained targets, INT8 post-training quantization and QAT, and deployment across TensorRT, SNPE, QNN, CVFlow, NCNN, MNN, and TFLite.
I hold three patents filed at VinAI covering gaze prediction, face recognition training, and driver monitoring on edge devices.
Academic service — Conference reviewer for WACV 2024, ECCV 2024, CVPR 2025, BMVC 2026, and NeurIPS 2026.
Interests
- Generative Models & Diffusion
- Face Analysis & Perception
- Vision-Language Models
- Efficient & On-Device Inference
Education
BSc in Computer Science, Honor Program
Posts and Telecommunications Institute of Technology, Viet Nam
6+
Years shipping CV & generative AI
10+
Edge targets deployed to
3
Patents filed at VinAI
Stack
PyTorch TensorRT SNPE QNN CVFlow TFLite NCNN MNN LibTorch C++ CUDA Diffusion VLM INT8 / QATSelected Papers
SEMACS: Semantic–Acoustic Speech Coding for Ultra-Low Token Rates
Under review, 2026
Joint semantic and acoustic speech coding aimed at ultra-low token rates. Under review.
Driver Attention Tracking and Analysis
arXiv preprint, 2024
Estimating where a driver is actually looking in the road scene, using two ordinary cameras instead of eye-tracking glasses — plus a large in-situ driving dataset with point-of-gaze annotations.
QC-StyleGAN: Quality Controllable Image Generation and Manipulation
NeurIPS 2022
A StyleGAN variant whose output quality is an explicit, continuous control code — letting you edit real low-quality photos without first “fixing” them, and getting blind image restoration for free.
Selected Patents
Driver Monitor System On Edge Device
Toan Bui Duc, De Vu, Dat Viet Thanh Nguyen, Tuong Nguyen, Tin Nguyen Truong Trung, Hung Bui
Method And System For Training A Machine Learning Model For Face Recognition
Duc Toan Bui, Van Tuong Nguyen, The De Vu, Viet Thanh Dat Nguyen, Hai Hung Bui
Method And System For Training A Machine Learning Model For Point Of Gaze Prediction
Dat Viet Thanh Nguyen, Anh Tran, Cuong Pham, Hung Bui
Selected Projects
Generative AI
AI Photo Editing App
One-tap photo edits — object removal, background removal, sky replacement, and outfit changing — built on generative inpainting and segmentation.
Sep 2023 – Jan 2024
Segmentation
Real-Time Portrait Segmentation on Mobile
A lightweight portrait segmentation network for phone cameras, plus the full data pipeline behind it.
Aug 2021 – Dec 2021
Object Detection
Industrial Surface Defect Detection
Detecting and grading surface defects on a production line, delivered as a .NET-integrable component rather than a Python service.
Jan 2021 – Apr 2021
Autonomous Driving
Autopilot Perception on Jetson TX2
The perception stack for an autonomous driving entry — lane, sign, and obstacle understanding running on an embedded Jetson TX2.
Jan 2020 – Jan 2021
Writing