University of Information Technology VNUHCM
Bachelor of Science, Computer Science
- Current GPA: 8.43/10
- Coursework: Artificial Intelligence, Computer Vision, Deep Learning, Data Mining
Tran Quang Khai
Machine Learning / Computer Vision
With a strong foundation in Machine Learning, Deep Learning, and Computer Vision, I aim to pursue graduate studies to deepen my theoretical understanding and learn to deploy and maintain machine learning systems in production. I am particularly interested in building automated data and machine learning pipelines.
Summary
I focus on building practical AI systems with an emphasis on computer vision, model deployment, and production-oriented pipelines. My recent work spans aesthetic-aware image cropping, real-time detection systems, and ML-driven automation across research and applied engineering settings.
Education
Experience
Lab Assistant
Multimedia Communications Laboratory
- Conduct research on aesthetic-aware image cropping using Large Language Models such as GPT and Gemini, contributing to a forthcoming publication in 2025.
- Developed and published two datasets for aesthetic-aware image cropping, evaluated on those datasets to establish a benchmark.
Technical Specialist
Vietnam-Singapore Industry 4.0 Innovation Center, Eastern International University
- Participated in developing AI for HVAC systems to control the school's AC, achieving fan speed accuracy of 91%.
- Integrated the YOLO11x model for people counting on the CCTV server for building management using DeepStream, achieving real-time human detection on multiple cameras.
Projects
Aesthetic-Aware Image Cropping
- Project description: Build a pipeline for prompt-based image cropping using aesthetic-awareness.
-
Methods:
- Generative AI models like GPT and Stable Diffusion to generate synthetic datasets.
- Leveraged models such as CACNet, SAM2, and Grounding DINO for image cropping.
-
Results:
- Published two datasets: 1,000 manually labeled images and 4,000 automatically labeled.
- mDSC: 0.522 using Gemini-2.0-Flash, and 0.453 using GPT-4o mini.
- Tools: Python, GPT, Stable Diffusion, TensorFlow, Gemini, SAM2, Grounding DINO.
Building a Real-Time Human Detection System
- Project description: Using Nvidia DeepStream for real-time human detection on CCTV.
-
Methods:
- Fine-tuning YOLO11x for maximum accuracy on human detection.
- Converted the model into TensorRT to boost FPS and apply batch processing.
- Integrated MediaMTX to transcode RTSP into low-latency WebRTC.
-
Results:
- Optimized YOLO11x inference using TensorRT FP16 precision, enabling the system to scale from 1 to 10 concurrent video feeds on a single GPU while maintaining 20 FPS real-time.
- Tools: Python, DeepStream, Mediamtx, Caddy.
Skills and Technologies
Languages: C++, Python.
AI & Computer Vision: NVIDIA DeepStream, TensorRT, PyTorch, OpenCV, Ultralytics YOLO, CUDA.
DevOps & Deployment: Docker, Docker Compose, Git, Linux (Ubuntu/WSL).
Cloud Platforms: Azure (familiar with compute and basic networking).
IELTS: 8.0 (L: 8.5, R: 9.0, W: 7.0, S: 6.5).