Recommender Systems
Personalized systems
2017-21 | Engineering, Industrial R&D 🔗
Worked as a Tech lead in AI and Data Science, which was essentially a "full-stack" AI Engineer/Type-B Data scientist who analyzed business requirements, researched and prototyped, developed, productionized and coordinated to deliver AI/ML-based recommendation solutions for, but not limited to, e-Commerce and Digital Media corporate customers. Machine learnt algorithms have been deployed and served in real-time to 6+ million end-users in the last 12 months, increasing user engagement by upto 25%.
A Bayesian model for Content-based Collaborative Filtering recommendation
2016-17 | Research | Published
🔗
Designed a novel, scalable, and interpretable Bayesian model which included textual description of the items for
recommendation. The algorithm was evaluated on both public and real-life, commercial datasets in different
recommendation settings, namely article recommendaiton and product recommendation. It showed noticeable
improvement in prediction accuracy against other competitors with fast training time on CPU - thanks to
optimization on convex-hull of the loss function.
The work was published in IJAR - an ISI journal, ranked Q1/Q2 in Artificial
Intelligence and Applied Mathematics - and adopted by an industrial partner.
Visual computing
Artistic Photo Stylization
2017 | Indie dev
Developed server-side image processing service of a casual image editing app on iOS, based on (deep) Style transfer - a
collection of modern advances in data-driven, learning-based approach to Image analogy and Optimization.
Multi-label Image classifcation with no Localization
2016 | Research | Published
🔗 | Code 🔗
An early work that explored ConvNet for image tagging with single-scale input. The single-scale approach was
different from e.g. R-CNN, YOLO framework in which multiple labels can be classified without extracting
candidate regions before prediction, thus computational complexity is much lower at test time. Experiments on 2
small-scale datasets showed competitve classification accuracy, though fell short in
rank-wise retrieval metrics - which was believed to be amendable by training against pair-wise loss functions.
Though a humble work by today standard, this was one of the 3 papers that kick-started Deep learning research at Data Science Lab, HUST. At personal level, it was my first of many: 1st academic publication - also 1st Best paper
award, 1st non-trivial HPC work, and 1st time being a thesis supervisor.
Graphical Overhaul
2007-08 | Game mod, 3D Modeling
Modeled and ported higher-poly 3D models into a retro 90's game to upgrade the game visual.
The ports were possible thanks to community-developed tools that
helped import/export game assets.
Original (1997)
Mod'ed (2007)
All game characters are © by SQUARE ENIX / SQUARESOFT