Showcase the outcomes of selected projects I implemented in the industry, which I find interesting
This project focused on improving the Click-through Rate (CTR) predictor at Knorex, a core element in optimizing ad campaign performance for advertisers. By utilizing model calibration and continual learning, I enhanced the accuracy and adaptability of the CTR prediction system. Due to the confidential nature of the work, specific details of the implementation will not be discussed. This result achieved a significant improvement in revenue and performance of the platform, which was highlighted on the official company website.
Introduction In this project, I developed a system for tracking multiple individuals across multiple camera views, both overlapping and non-overlapping. The aim was to create a reliable solution for real-time tracking of people, which could be applied in various scenarios like public safety and surveillance. This showcase highlights the key results and methods used, focusing on the overall outcomes and the system’s capability, rather than the specific implementation details.
Some of my personal/student projects and hobbies
I could have chosen to learn badminton from a tutor, but I genuinely enjoy figuring things out by myself. Three years ago, I was a complete beginner in the sport, just like anyone starting something new. I watched countless tutorials and self-taught the fundamentals. I focused on key aspects like basic swings, body mechanics to efficiently create a powerful-yet-natural movement using my wrist, arm, shoulder, and hips. I took a bottom-up approach, mastering one element at a time, and prioritizing consistency over speed. The most crucial part of my self-learning journey was feedback. I started recording my sessions, analyzing my movements, and comparing them with pro players to ensure I was on track. Even after sustaining injuries—once with wrist arthritis and multiple times with ankle sprains—my focus on correcting posture and movement only grew stronger. This journey led me to create this YouTube channel, where I documented and shared my progress over the last three years (quite a few views, but I was not creating this for any fame-related purpose in the first place 😁 I usually put most of them in “Unlisted” so that only those who have the URLs can view). Contact at the right height? Moreover, I even discovered an app on the Apple Watch utilzing its built-in gyroscope and accelerometer sensor to detect tennis actions and measure swings metrics during a match. I immediately thought it could adapt to badminton because their nature are similar to some extend. However, the accuracy turned out to be average on badminton-specific actions. Unstoppably, I attempted to retrieve the raw data from the Apple Watch to build my own ML model tailored to badminton, but I hit a roadblock as the data was partially censored for the level of analysis I needed and I had to stop. Smash Swing Speed Backhand Swing Speed Having said that, the process of recording and reviewing not only helped me improve, but also brought joy and lasting memories. It allowed me and my friends to analyze our games, discuss what made a point good or bad, and foster a positive learning environment where we could all improve together. Today, I have to pause this hobby with my friends as I’m flying to another country to pursue my studies. Although I’m stepping away from badminton temporarily, this self-learning journey has reinforced how I approach other challenges in my life.
I contributed to the translation project of the popular “Dive into Deep Learning” book into Vietnamese. This project involved translating complex machine learning concepts and examples into clear and accurate Vietnamese, making this important resource accessible to a wider audience. My work ensured that the technical accuracy of the content was preserved while also tailoring the language for Vietnamese learners, helping to bridge the gap between global research and local development. Link to the book: Original version, Vietnamese version Link to the project: https://github.com/d2l-ai/d2l-vi