I want to build things that can generate tremendous and long lasting values for the society. These are the things that won’t expire or go outdated quickly, things whose value can grow and accumulate over a long period of time, and things that can continuously generate positive values for people and the society.
During my undergraduate study, my major was industrial engineering. Through my program, I learned about people’s relentless efforts for optimization in the manufacturing process, and many innovations that triggered significant impacts on the landscape of manufacturing and have continuously created values till now. For example, the Scientific management theory by Federick Taylor and Fordism by Henry Ford were both developed in 1880s and they started the a revolution for industrialized, standardized mass manufacturing process. This line of thinking kept evolving Toyota Quality System and Lean Six Sigma. They largely shapes the modern manufacturing industry, fostered the post-war economic boom in US, and bring the productivity of factories to an incredible level, which now allows us enjoy massive-produced, high quality products like iPhones and cars. That was when I got inspiration to build such innovation that can profoundly change an industry and bring meaningful impacts to human lives that would last over a century.
After I graduated from University of Toronto, I was fortunate to join Prof Andrew Ng’s startup Landing AI, a venture where I can potentially create profound and enduring impact to the manufacturing with the latest AI technology. Here we work with manufacturers to automate their quality inspection process with machine learning. Nowadays, majority of factories are still relying on human workers to carefully examine manufacturing products and check for any defect. This task is very repetitive but challenging. If the human workers made a mistake, which happen very often, the manufacturers would face large cost of recalling defective products and receive significant damage to their reputation. One customer that I have worked with is a global compressor manufacturer based in Brazil. Before my team started at the project, the manufacturer relied entirely on human taking 3 shifts a day to examine every compressors from the assembly line to catch any defective compressor with air leakage and send them to repair. We started by designing an imagery solution from scratch and quickly iterate on the camera sensor, lens, and configurations like aperture, shutter speed, and ISO. From the input videos of 5 megapixel resolution, the object of our interest can be as small as 4 pixels and moving at 30 cm/s at noisy background. To capture the defects, we built an end-to-end software pipeline that includes four different machine learning models for semantic segmentation, object detection, blob detection, and classification. Furthermore, we deployed this automation system into edge-computing devices and integrated seamlessly into a factory in Brazil. Each device streamed the camera feed to the machine learning system and made real-time inference to support the high production throughput in the factory non-stop for 24/7. Running in production for months, the system quickly matched human performance in precision and recall and it sometimes captured defective products that human completely missed. To ensure successful deployment, we solved numerous challenges, including standardized data labeling, compressing our models for limited device memory and latency constraints, and optimizing an imagery system for detection of very small but rapidly moving objects. The system generated tangible values to the customer’s manufacturing process: the overall inspection quality got improved and the workers who used to work at this repetitive inspection process are free to work at tasks that are more interesting to them. It triggered lots of meaningful discussions from customers and they plan to expand it to all their factory sites globally.
While we went through such project one after another, we reflected from our experience and iteratively developed a systematic machine learning life cycle for building an effectively AI system for manufacturing. We gradually discovered this system could be useful to many manufacturers across domains and all over the world. In the beginning of this year, I visited a steel manufacturer in China. They are one of the biggest steel makers in the world and they are actively pursuing the automation of quality inspection in production. They invited multiple world-leading AI companies to participate in their challenge. In the challenge, they provided us a proprietary dataset with around 40 defect classes. All the defects are specific to steel production and there are lots of ambiguities between classes. The problem was so difficult that the internal research team in the company iterated a visual system for almost a decade and achieved just 90% precision; other competing teams spent a month and their models reached 85% at best. I went there alone and hired three interns from local to help me work on the problem. I went through our system to prepared the dataset, sat down with their quality engineers to clean the labeled data, iterated machine learning models, and conducted error analysis. In two weeks, my system achieved over 93% precision and run at 30 frames per second, outperforming other teams by a significant margin. That helped Landing AI won the first place at the competition, and that was a moment where I believed we are getting it correct at the automated visual inspection. Soon after this trip, at Landing AI we started to build a platform that supports end-to-end development of machine learning system for quality inspection and we test it with more manufacturing customers. It is the LandingLens that Landing AI officially announced to the world in October 2020. I am very proud of this work for not only we develop a very successful approach to solve big challenge at quality inspection, but also we are democratizing it to all manufacturers all over the world to help them develop and deploy such AI system at their factories. This is what I believed that can generate tremendous and lasting values for people at manufacturing over a long term.
My undergraduate study in industrial engineering at the University of Toronto provided me with inspirations to build an innovation that can profoundly change an industry. Throughout my work at Landing AI, I am more confident than ever that I want to continue to pursue my current path of building successful AI products to build automation and improve efficiency for all kinds of industries. Going forward, my main goal is to learn how to build scalable AI products that enable machine intelligence in a broad range of industries, and thus create enormous value for society over a long term.
This was a part of my statement of purpose for grad school application. After I wrote it, I felt I can share it out publicly to keep remind myself about my goal.