Over the past four and a half years at Landing AI, I have had the incredible opportunity to work with Andrew Ng, Dillon Laird and other amazing people to build AI applications across various industries. Each project has brought its unique challenges, pushing me to dive deeper into the ever-evolving world of AI. As I look back at this enriching journey, I am grateful and humble to share the lessons that I've learned in the hope of inspire others in the field.

1. The Power of Computer Vision (CV) to Revolutionize Industries

I love computer vision in many ways. This passion, together with the mission to democratize AI for the public, led me to join Landing AI at its founding. Computer vision bridges the world of atoms with the world of bits, enabling innovative solutions that have a lasting impact on the physical world.

During my time at Landing AI, I developed computer vision applications to tackle a wide range of real-world problems. We went into factory facilities to deploy automated inspection solutions to improve product quality (manufacturing), installed cameras and GPUs into harvesters to optimize wheat harvesting in the field (agriculture), and created intelligent tools for analyzing battery production in electric vehicles (energy). The possibilities are endless. Each project revealed the tremendous potential of computer vision applications to revolutionize industries, improve efficiency and productivity, make jobs easier and safer, and drive concrete business value.

In the past two years, we have been developing a cloud-based MLOps platform to enable everyone to build such applications with ease and harness the potential of computer vision systems (see this blog). I believe this is just the beginning. With the emergence of Foundation Models (like Segment Anything), the barrier to building successful CV applications will be even lower. In the following years, we may see proliferation of simple yet transformative computer vision tools across various industry sectors and improve our daily lives.

2. The Importance of Data Curation at ML

Having worked on dozens of real-world AI projects, I have learned from the bottom of my heart that the quality of data used to train and evaluate machine learning models is incredibly essential in the recipe for success. At Landing AI, we have realized that rigorous and meticulous data curation is paramount in unlocking the full potential of the models we create. This realization led to the Data Centric AI campaign launched by Andrew Ng in 2021, which was highly echoed and embraced by the community, and has become the underlying design principle of our platform.

Data curation comprises several facets, including data collection, active sampling, label quality management, data cleaning, augmentation, etc. For real-world AI applications, from cat detector to autonomous driving and ChatGPT, significant system and engineering efforts are invested in improving data quality.

Labeling quality management is especially critical in applications developed at Landing AI. The problems often involve non-common objects and non-trivial labeling rules. Confusion among class definitions and conflicting labels are common root causes for low model performance. To address this, we spent considerable time sitting down with subject matter experts to correct labeling misalignment and clean labels. It was surprisingly productive exercise, yet undeniably tedious and painful. Ultimately, this led to more systematic and automated features, such as automatic mislabel detection.

Label quality is only one puzzle of the entire picture. Each of these aspects in data curation is crucial to achieve successful production-ready AI systems. I am looking forward to explore more on active sampling and data generation at scale in the future.

3. Improving Iteration Speed in Experiments

Andrew Ng once shared with me that when he recorded videos for the machine learning course, he did so in two-minute chunks. After each recording, he watched the video to identify areas for improvement. This process allowed him to iterate more quickly. At Landing AI, we have adopted this method to improve our own iteration speed.

I recall in the early days, I would launch training jobs before going to bed, and then check the results upon waking up in the morning. This was because jobs took several hours to complete and I didn’t want to waste my sleep hours. This iteration speed slowed us down and incurred large training costs, so we decided to make some changes. In one year, with a collateral efforts from a group of talented engineers, we were able to reduce training time from hours to just five minutes! This faster experimentation cycle fundamentally changed the way we conduct experiments. We can now rapidly test treatments (usually on data side), validate hypotheses, and receive instant feedback, all achieved with less training costs.

ML engineering is an inherently iterative process. Faster experiment cycles lead to quicker progress and better results in projects. I am confident that this principle will be carried on with me now and in the future.

4. The Secrets of Making AI Products

Developing successful AI products demands more than just technical expertise. At Landing AI, I have come to realize through hard experiences that comprehending user needs, designing simple and intuitive experiences, iterating based on user feedback, and evaluating using the appropriate metrics are all essential components of creating products that people will be willing to try and, hopefully, eventually love.

The concept for our MLOps platform was first conceived in 2020, but it underwent numerous iterations and two major rounds of complete redesign before reaching its present state in 2023. In the first design, we designed every step so that it can follow our proposed ML life cycles and best practices. However, the results were disastrous. Users loathed our workflow and found it difficult to follow and complete projects. We learned to listen to users' concerns, adapt to their requirements, yet invent novel techniques and experiences to alleviate their pain points more effectively. While I wouldn't say we have achieved perfection yet, I am proud to present this product to anyone who might find it beneficial.

As the platform continued to evolve, our focus is on incorporating more advanced features while maintaining a user-friendly interface. We start with users’ pain points and develop prototype with real examples that reflect them. Once the prototype is ready, we begin to involve early users in the development of MVP, went through the prototype with them, observe their reactions, and listen to their feedbacks. Based on that, we iterated and developed further. Upon release, we monitor metrics to ensure that the features are performing well and that users enjoy using them. It is important to maintain an ongoing feedback loop with end-users throughout the product life cycle. This ensures that product enhancements are continuously informed by user experiences, ultimately resulting in a more successful and user-centric AI solution.

5. Thriving Through Startup Cycles

Perhaps one of the most significant lessons I've learned at Landing AI is the importance of perseverance and resilience in a startup environment. Startups are inherently characterized by ups and downs, and one must ride out the cycles and learn from each experience to emerge victorious.

During my time at Landing AI, there were moments when I felt extremely stressed and lost sleep. There were also moments when we made pivotal changes in the company's direction, and the future seemed uncertain. There were even moments when success seemed impossible to achieve. These were all common yet challenging times in startups. However, I am glad that I persevered through those times. It was the people around me who gave me confidence and support to surmount obstacles. Through those times, I built friendships and achieved personal growth.

If you asked me what it's like to be at this AI startup, I'd say it's an experience of deep dives into new technologies and constant discussions about how to create value with them. It's challenging and pushes me to expand my skill set beyond research and engineering, and into product and business. Most importantly, I'm surrounded by innovative, brave, and caring people who help me grow as a person.

Final Words

Without a doubt, my journey with Landing AI has been tremendously rewarding. As I reflect on my 4.5 years with the company, the most prominent theme would be "learning." I vividly remember my first few days at the company, feeling excited and thinking, "I breathe AI day and night here." Since then, I have worked on numerous projects and gone through different stages of the company. What remains constant is my eagerness and exposure to learning — learning about computer vision, MLOps, product development, sales, and more. I have thoroughly enjoyed my learning and growth here.

With my goal of continued learning in mind, I have decided to move on and embark on a new journey. My passion for computer vision and its real-world applications has been the driving force behind my latest decision. My focus will be on exploring multi-modality and foundation models, as I believe that these areas have the potential to unlock new and exciting possibilities in the field of computer vision. I am excited to take on this new chapter of my life and can't wait to see where this journey takes me.

I would like to express my gratitude to Landing AI for the valuable experience I have gained during my time here. I wish Landing AI success in their mission to help more people embrace AI technology. The work that Landing AI is doing to make AI more accessible to everyone is truly inspiring, and I am honored to have been a part of it. I will always cherish the memories and skills I have gained during my time at Landing AI, and I look forward to carrying these experiences with me as I embark on this new chapter in my life.