May 2016 - Aug 2016
I interned at a start-up company named GoFind AI during the summer of 2016 and it was an incredible experience. GoFind AI is a company that develops a visual search engine for e-commerce, enabling customers to upload an image of the product they desire and retrieve results of incredibly similar items from its database. During the 10-week internship, my job was divided into three phases. Phase one was scraping data from multiple commercial website. Phase two was about implementing the K-Means algorithm using Apache Spark and applying it on AWS X1 instances to process and categorize over 25 million images. Phase three was using the Tensorflow framework and inception model to train classifiers and extract features from images based on styles.
Phase three was my favorite. In detail, it was about creating a multi-class image ranking system. The service consisted of multiple machine learning models organized into a two-level hierarchy. The root was a classifier that recognized the product and supported 40 different products. The second level consisted of one ranking model for each image category.
The project was then divided into two sub-phases. During the alpha phase, my team and I developed the full stack, including the feature extractor, classifier, KNN rankers, integration backend code, database storage, AWS infrastructure as well as mobile client. The purpose of the trial run was to develop all the components and integrate them in a working system.
During the beta phase, we attempted to fine-tune the inception classifier with 300 categories and 5000 images per category. For this we would have to modify the inception and run the fine-tuning on the IBM cloud multi-GPU instance (since it was a cheaper and more powerful option than AWS). After all the steps were done, we developed a REST API to store the data into the MongoDB.
GoFind gave me a chance to see how early-stage startup company operated. I was excited to see ideas turning into products.
Please download our iOS App!