I am currently a principal engineer at Azure GenAI. My main focus is to bring GenAI to practically serve Microsoft and its custmers on realistic scenarios. I help to drive the creation and release of the Phi-3 model family including Phi-3.0 (May and June 2024), Phi-3.5 (August 2024) to both open source community and Azure AI. Previously, I drove AI strategies at scale and bring clarity to the leadership team at the Microsoft Office of the CTO.
My professional career is a mix of industrial research labs and startups. I spend a few years at the Machine Intelligence Technology Lab, DAMO Academy, Alibaba. My main focus is to break the language barriers across the Alibaba ecosystem by researching and developing AI solutions for eCommerce scenarios. I spent 2 years in Textio, a start-up of augmented writing, where I was responsible for machine learning models. I worked for Microsoft on machine learning models in wearable devices such as the HoloLens project. I was a machine translations researcher at SDL. I am actively coaching and consultanting early-stage startups and young engineers in Vietnam.
Specialties: GenAI, AI strategies, natural language processing, machine translation, speech translation, multi-modal AI, confidence estimation, language modeling.
PhD in Language Technology, 2012
Carnegie Mellon University
MS in Computer Science, 2005
Johns Hopkins University
BSc in Maths & CS, 2001
Vietnam National University, Hanoi
As the 1st machine learning engineer, I’ve helped build Textio’s core predictive engine and learning loop for the augmented writing platform which already used by thousands of companies worldwide.
Spearheaded the development of the Textio core models with cutting-edge technologies in statistical natural language processing and machine learning.
Design, develop, ship, and improve production features, such as prediction engines for equal opportunity employment, job type, and document type.
Created scoring models that helped increase predictive power significantly while preserving explainability and interpretability.
Working on the next generation of wearable devices at Microsoft, e.g. HoloLens:
BCI with deep learning models, e.g. CNN, LSTM, GRU, with a patent pending on eye tracking technology.
Implement speaker verification systems on DSP which includes enrollment with MAP adaptation, verification with novel scoring methods, and back-end training pipeline for GMMs.
Reduce memory footprint and speed up runtime for i-Vector speaker recognition system with matrix factorization. Implement average stochastic gradient descent with L2 regularization to train sub-matrices.
Research on deep neural network for brain computer interface, i-Vector, probabilistic linear discriminant analysis, matrix factorization, and DNN for multiple-speaker identification.
R&D in commercial machine translation systems.
Model adaptation: worked on techniques to automatically adapt background translation system to a specific domain/genre via information retrieval approach and machine learning methods.
Confidence estimation: explored methods for machine translation quality-prediction including SVM and M5P decision tree. Member of the SDL Language Weaver team that won the 2012 MT quality prediction competition.
Reordering models: implemented lexicalized reordering models with distributed Hadoop/Pig training pipeline and real-time decoding.
Guiding teams to reach ambitious goals
Ensure seamless teamwork.
Enough to get things done timely
Be able to explain LLMs to a kid
Extract gold from dirt
Turn research ideas to business opportunities