AI Researcher

Technical Skills

• Programming: Python, SQL, Matlab, C/C++
• Machine Learning & AI: PyTorch, OpenCV, YOLO, UNet, Neural Network, CNN-LSTM
• Database Management: SQL Server, Stored Procedures
• Cloud & Deployment: Docker, NVIDIA Jetson (TensorRT), AWS
• Tools & Platforms: Git, Linux (Ubuntu), Jupyter Notebook
• Project Management: Agile (Scrum), Technical Writing, Publication

Education

• B.S., Computer Science and Engineering

  Vietnamese German Univeristy, Vietnam (Juli 2023)

Work Experience

AI Researcher @ Vietnamese German University (January 2023 - Current)

• Conduct comprehensive literature reviews to identify and implement state-of-the-art ML methods for real-world applications, directly aligning with the role’s requirement for core ML expertise.
• Design and implement data collection pipelines, preparing large-scale datasets through cleaning, normalization, and feature engineering to enable robust model training.
• Collaborate cross-functionally with research teams on model development, demonstrating strong teamwork capabilities mentioned in the job requirements.
• Build validation frameworks for ML models to ensure continuous performance improvement, similar to the A/B testing mentioned in the job description.

Data Engineer @ FPT Software Company (Feburary 2022 - January 2023)

• Implement SQL Server database architecture following Software Requirement Specification (SRS) and Mockup from Business analysis
• Developed stored procedures for data retrieval, cleaning, transforming, and logging between databases and web applications.
• Used Scrum for Agile project management to collaborate with cross-functional teams.

Projects

Neural parameters calibration for Dengue transmission modeling with real epidemic dataset from Colombia, Peru, and Puerto Rico

(On Progress)

• Developed and deployed Neural Network (NN) with multi-agents system to calibrate parameters for Ordinary Differential Equations (ODE) dengue modeling.
• NN shows competitive results but 25x faster regarding the speed, compared to tradition Markov Chain Monte Carlo (MCMC) traditional calibration.
• My code is developed based on open-source NeuralABM (Github). I’ve also added many techniques such as learning rate scheduling, gradient clipping, dropout, CNN-LSTM architecture with batchnorm, to the base code to improve the performance of Neural Network and the experiments.

Key technologies: Neural Network, ODE, multi-agents, CNN-LSTM

Benchmarking Jetson Edge Devices with an End-to-end Video-based Anomaly Detection System

(Publication) (Github)

• Developed and deployed an AI-based anomaly detection system on Jetson edge devices, achieving 84.39% AUC on UCF-Crime and 88.4% AUC on UIT_VNAnomaly datasets.
• Optimized deep learning deployment with TensorRT and Docker, fine-tuning models to achieve 47.56 FPS with 3.11 GB RAM usage on edge devices.
• Benchmarked deep learning models on NVIDIA Jetson platforms, improving performance by 15% while reducing energy consumption by 50%.

Key technologies: RTFM (video-based anomaly detection), slowfast, Jetson Edge devices, torch tensorRT

Water Body Detection Using UAV Multispectral Imagery

(Under Review) (Github)

• Collect and construct the high-quality WaterMAI dataset (880 images), optimized for small and medium-sized UAV-based water body detection.
• Evaluated deep learning models UNet, MSNet, Yolov7, Yolov7x using bounding box and segmentation approaches, achieving Dice scores > 0.92
• Optimized spectral input combinations (RGB, NIR, NDWI) to enhance model accuracy for environmental monitoring.

Key technologies: UNet (segment), MSNet (segment), Yolov7 (bbox), Active Learning (for segment/bbox data labeling), multi-spectral processing

Self-driving car simulation using behaviour cloning

(Github)

• Use ROS 1 Noetic, Gazebo simulator for demo and testing the self-driving algorithm
• Autonomously drive on the designated map (highway, curves, intersection)
• Apply “Behavioral clonning”, taking input from camera and output steering angle and speed

Key technologies: ROS 1 Noetic, Gazebo, behaviour cloning

PM2.5 Concentration Prediction Using Traffic Surveillance

(Publication)

• Benchmarked object detection models YOLOv5, SSD, Mask-RCNN for vehicle counting to estimate PM2.5 level from specific formula.
• The final result shows 0.9 Pearson correlation between traffic density and PM2.5 levels.

Key technologies: Yolov5, SSD, Mask_RCNN, data labeling, Kalman Filter (tracking algo)

Publications

1. Pham, H.V., Tran, T.G., Le, C.D., Le, A.D., Vo, H.B. (2024). Benchmarking Jetson Edge Devices with an End-to-End Video-Based Anomaly Detection System. In: Arai, K. (eds) Advances in Information and Communication. FICC 2024. Lecture Notes in Networks and Systems, vol 920. Springer, Cham. https://doi.org/10.1007/978-3-031-53963-3_25
2. C. D. Le, H. V. Pham, D. A. Pham, A. D. Le and H. B. Vo, “A PM2.5 concentration prediction framework with vehicle tracking system: From cause to effect,” 2022 RIVF International Conference on Computing and Communication Technologies (RIVF), Ho Chi Minh City, Vietnam, 2022, pp. 714-719, doi: 10.1109/RIVF55975.2022.10013864.
3. (UNDER REVIEW) Ngo Phuc Linh, Pham Viet Hoang, Bui Ngoc Long, Vo Bich Hien, Thirumalaisamy P. Velavan, Tran Duc Khanh. Use of low-cost drones to map arbovirus vector habitats with multispectral aerial imagery, 14 February 2025
4. (ON PROGRESS) Pham Viet Hoang, Nguyen Trung Dang Khuong, Thirumalaisamy P. Velavan, Tran Duc Khanh. Neural parameters calibration for Dengue Outbreak Forecasting, 08 April 2025