 Hello, everyone. This is Lin Du, the founder and the CEO of Basic AI. Today, I'm introducing Xtreme One, the next generation open-source multi-model training data platform. A short introduction about Basic AI. I founded Basic AI seven years ago, aiming to speed up the evolution of the AI landscape and to build a more intelligent world together. Basic AI is globally inclusive. We can find us in Erwan, California, Seoul, Korea, and Chengdu, China. So during the past seven years, we have accumulated more than 10,000 different projects and are serving for hundreds of the clients all over the world. And also we presented the data-centric ML labs, especially Xtreme One, and the first open-source multi-model training data platform. And this product is built by the TechSuite team, which has very strong machine learning foundation and strong belief in open-source. And all the engineers graduated from top universities or worked in prestigious companies before. And also Xtreme One is a data-driven product. We have accumulated more than 100 terabytes data, and we have also abstracted the key pay points from the processing of those data. And our product is trusted by thousands of global partners, including Stanford University, Intel, Berkeley, etc. Xtreme One is built to accelerate the modeling process with advanced AI power tools. And also we come with entomologists, which were distilled from thousands of the projects. And also we offer a wide variety of data-curing features. So Xtreme One went open-source since September of the year 2022. And three months later, Xtreme One became a hosted project on the Linux Foundation AI and data. And so far, we are still the only project in the category of data annotation of the landscape by Linux Foundation AI and data. So why we build Xtreme One? UBS Global Research reports that now AI engineers are spending 70 to 90% of their time on training data. So from our past experience, we see that data has become the new bottleneck in AI model training process. Especially engineers are spending too much time on defining the data annotation schema to clean the data and to define the data pipeline. So we see there are lots of the repetition in data and the model fine-tune process. And now data is becoming more and more complex. From the previously single-frame image data to nowadays complex multi-modal training data, including the data from complicated sensors, LiDAR, radar, etc. And also engineers are facing data quality issues, garbage in, garbage out. So engineers are spending lots of time on improving the data quality. And also we are facing the data drift problem. So how to control this data flow to make sure our test data and training data are in coordinates? What is more, the data security problem? The companies, they are paying more attention on protection their data. And also the data labeling is very expensive as we all know. And all this labeling process lacks of the transparency. We have to face with the fragmented machine learning tools. So all these pinpoints we are addressing come to our solution. We call it the next generation training data platform. So speaking of the next generation, we are highlighting on the point that we would like to increase the efficiency of AI engineers. And for the past generation, most of this kind of training data platform, they focus on how to increase the efficiency of data labelers. The highlights of Extreme One. First, I would like to introduce Ontology Center. Ontology was the first introduced by Extreme One to abstract the definition of AI problems from the various model requirements. So actually Ontology is the abstract from the real-world problem to the data schema. And it can be reused and extended to build the knowledge base of AI algorithms, thus accelerating the model development. And from the Ontology Center, we can also do scenario search, recommendation, collaboration, and disambiguation. Also in Extreme One, we offer a full suite of annotation tools, including the most hard tool set to deal with a large language model, and also LIDAR basic and fusion tool sets, LIDAR segmentation tool sets, image and video tool sets, and task management, performance management, workflow management, and AI assisted annotation. Recently, the wide applications of large language model have skyrocketed and which also require a full tool set for reinforcement learning from human feedback. And in Extreme One, we come with the tool set, the full tool set for large language model, including the classification of the conversations from large language models, and also the rating, the dragging, etc. We also built 2D and 3D fusion annotation tools. That's why we call our platform as multi-model training data platform. In Extreme One's point cloud tool set, it is able to handle up to 150 million points count in continuous frame scenarios. And integrated auto segmentation models can also boost the efficiency by AI trainers. In Extreme One's annotation tools, we also support a full range of the model capabilities, and along with the parameter tuning capability. So the user may decide their own parameters during the model inference process. Here are some samples. In this slide-out fusion project, the user may customize their own parameters and the models. And one key, the results will be automatically calculated. And human laborers will only have to check the results and make minor changes of the result. Also, Extreme One provides very flexible workflow management for a team of all size with all quality requirements. The users may very flexibly set up the tasks and design the workflow and assign the corresponding tasks to different team members. The second most important part of Extreme One is its data curation feature. So actually, Extreme One is not a data labeling tool, but also it provides the full suite for data curation features. Like data visualization, versioning, organization, validation, consensus, augmentation, data drift analysis, and synthesis. For example, we support multi-model data visualization. And in this toolset, the user can easily check up all the results and to find the common cases to further address and to change. Extreme One also incorporated with Sata models. So actually, it is an AI-powered platform for maximizing the efficiency of data and the model productions. We use transfer learning to do the pre-train models. And we offer easy application for these Sata models. So actually, it's like we provide the environment for fast modeling and lifelong continual learning process. The last part is Extreme One's techniques. As an open source software, Extreme One also builds upon different open source solutions. So we compliance with the principles of cognitive architecture to ensure the scalability, elasticity, stability of the platform services. We have built an entire architecture, which is composed of an access layer, application service layer, base service layer, runtime abstraction layer, and infrastructure layer. Extreme One, the multi-model training data platform, faces challenges in managing and saving large-scale data in various formats, such as images, LiDAR point clouds, video and NLP data. To address these challenges, we rely on different technical solutions for structured and unstructured data storage. For structured data storage, we utilize TIDB, which offers a flexible and distributed data architecture that supports high availability. It is also compatible with MySQL and provides a rich ecosystem of tools for deployment, transfer, and synchronization. For unstructured data storage, we employ MinniO as a solution, which offers an efficient storage plan for labeling and accessing multi-sensory data. MinniO is compatible with the S3 API and provides features such as data loss protection, server-side encryption, and data persistency after writing. Extreme One also provides a unified scheduler for both multi-cloud and heterogeneous clusters, accommodating deployment in public and private clouds, to manage development, testing, and production processes. We use Kubernetes together with Ranger, which offers a user-friendly interface for quick application deployments and supports GPU sharing, where a media's Kubernetes device plug-in. In a sync computing process, we address challenges like data preparation, massive data sets rendering, compression, and chunking. To tackle these issues, we rely on Pulsar for cloud-native support, isolated computing and storage, and high performance with low latency. Extreme One originated from open source, and embraces the open source. So we would like to build a strong-native community with a culture of helping each other and make this product better and better, because we support a full lifecycle of data-centric MLOps. Incubated by Linux Foundation and data, Extreme One is experiencing fast growth. On the landscape of MLOps, Extreme One sits in the labeling category. We are looking forward for more collaborations with the open source community. Now we have set very clear roadmap for Extreme One. We will be enhancing the capabilities of Extreme One in more aspects, mainly in data versioning, visualization, and organization features. Just in last week, we had a new milestone, a thousand downloads from Docker download platform, and also we have 360 stars so far. We are still growing on a fast pace. Today, Extreme One has been used in multiple industries, including autonomous driving, smart retail, smart home, smart city, etc. We are looking forward to more applications of Extreme One. If you are interested, please visit Extreme One on GitHub. And welcome to join us and become part of our community. Thank you.