Developed the VoiceAI that was used for various platforms:
Mobile Conversational APPs, Car Voice Assistants for Mercedes and Hyundai, and Restaurants AI Voice Order System.
Particularly focus on badly-form conversational user voice queries voice.
Built and maintained a Audio-Text Command Query Noise Detector using Distil-RoBERTa (BERT)
transformer neural network, that could filter out 40% of noisy user voice-commands and resulted
in a reduction of 20% of processing time in the real-time query stream production pipelines.
Defined noisy queries for production feasibility purposes and user-centric purposes.
Automated a backend benchmarking analytics pipeline of production noisy queries.
Developed a Rule-Based NLP program for recognizing valid user-commands and created a Grammar-Detector together
to assist the auto-labeling of the noisy training dataset (in which only 0.5% were properly manually labeled)
by implementing entity extraction, parsing and Part-of-speech (POS) tagging using spaCy.
Created and maintained a pipeline that could automatically add more rule-based patterns
when new production queries comes by implementing spaCy.
Experimented in Bayesian semi-supervised classification
that can potentially correct the biased rule-based auto-labeling processes mentioned above.
Performed text mining of the user intentions and sentence structural patterns to detect
whether a user query that provides complex characteristics, such as a query that has multiple user intents
or a question in a query that provides multiple choices to compare between or choose from. Programmed a tool
to suggest proper text regeneration and responses.
Applied Quantitative Methods (AQM)
Data Analyst, Co-op
Cross-team with Best Buy Data Team to perform Twitter Data Analytics
Built a content-based spam tweets filtering system using Naive Bayesian classifier, that reached an accuracy of 90%.
Developed a sentiment analysis model to evaluate the satisfaction improvement of
customers after they were being responded to by Best Buy customer service on Twitter
using the support vector machines (SVM).
Successfully classified different topics for Best Buy tweets using the Latent Dirichlet Allocation (LDA).
Performed gender classifications of Twitter usernames and achieved an accuracy of 92%,
by implementing the semi-supervised character n-grams algorithm.
Academia Experience
Simon Fraser University
PhD Research Assistant & Teaching Assistant
In my PhD research experience, I was focusing understanding how a correlated fracture network evolved in a low-correlated system and in a high-correlated systems.
They depicted distinct characteristics:
For low-correlated systems, the breakage or generation of the elements (or bonds) in the network more randomly. The avalanche of fracture takes place in the mid-late stage.
For highly-correlated systems, the correlation of breakage starts very early in the fracture dynamics.
Built a random graph model that studied the random failure dynamics in polymer networks in statistical physics study purposes and successfully forecast real-world polymer failure times.
Established a kinetic Monte Carlo program in Python for simulating the fracture processes that follows a Markov chain.
I will continuously update my YouTube videos, which are my presentations of my studies in machine learning and discussions with my knowledgeable friends. These contents are focused on the field of NLP and specifically highlight transformer-based models like BERT, GPT, and T5.
However, recently I also gained knowledge and presented on the topic of reinforcement learning.