Data Scientist

4/2020 – 5/2021

– Container Recognition System: I built a logic component that handles the signals received from 6 recognition modules. The logic component must begin when a truck is involved in container activities and return the container identification numbers according to the truck.
– Golf Scoring System: I have developed an application for correcting the golf pose. The system gives score measure for each individual body parts(hip, shoulder, ...), and finally results a summary points.

Pytorch

4/2019 – 4/2020

– Research defect classification on vibration signal in industrial processes: I used Bi-direction Long Short Term Memory and Attention mechanism on Bearing Fault dataset, which allows the neural network to focused on failure signals.
– Anomaly Detection on welding image from manufacturing chain: Due to the lack of labeled defected images, I applied an unsupervised method, which was using Auto Encoder, to trained on normal data to model learn the normal distribution and returned abnormal score to determine defected data.
– Research a base-line machine learning pipeline: shortage the process of training and testing image classification tasks, which allows engineers to define a list of pre-processing functions and get the output results.
– Research anomaly detection and segmentation: I research a method named Unsupervised Anomaly Detection with Generative Adversarial Networks, AnoGan, which is a generative adversarial network trained on normal data and be able to segment defect regions.
– Reduce processing time for sound classification project: the current company code base is using the Genetic Algorithm, I modified the feature selection component and made it run in parallel. As a result of this improvement, the processing time reduces from 12 hours to 20 minutes on a 32-core machine.

Outsourcing, Python, Pytorch

8/2017 – 4/2019

– Social Data Enrichment: the collected social user data is missing useful information such as
age range, location,... I analyzed the current information and came up with a hardly feature
engineering algorithm for data enrichment, which returned almost 90% accuracy.
– Chatbot System: Mobile Phone Shop Chatbot allows customers to ask questions such as
product features, shop locations, applications. I did mainly in implementation intent clas-
sification and named entities recognition, which taking advantage of Transfer Learning in
NLP, Universal Language Model Fine-tuning for Text Classification (ULMFit), to overcome
the problem of small dataset, and employed with Conditional Random Field for additional
constraints on nearby words to improve named entities recognition accuracy.

Natural Language Processing, Pytorch, Social Media Marketing