2018.3 ~ Now: CTO of AInnovation Limited & Chief Architect of the AI Instituteat Sinovation Ventures
Established in March 2018, AInnovation is dedicated to delivering AI-related products and business solutionsby leveraging the cutting-edge AI technologies, and empowering enterprise customers and partners to improvebusiness efficiency and realize digital transformation.AInnovation is committed to delivering “AI+” B2B enterprise services, with the brand influence of SinovationVentures, abundant AI resources, and support of AI experts from Sinovation Ventures AI Institute. Ourfounding team which is composed of elites from well-established companies such as Google, Microsoft, IBM,Baidu, SAP, Wanda, Huawei and Xiaomi, have an in-depth insight into the enterprise service market and haveaccumulated rich industry experiences and best practices. While focusing on the retail, insurance andmanufacturing industries, Our professional team works together with industry leaders and partners to helpstrengthen the competitiveness in the AI world.I am the CTO of AInnovation and leading the R&D team and responsible for designing and developing all ofour AI-related products and business solutions.
2015.12 ~ 2018.3: Chief Architect of Bigdata and AI at Baidu Cloud
Lead design/implement of all PaaS and SaaS products/solutions on Bigdata and AI at Baidu Cloud.
Baidu Ad pCTR ( Predict Click-Through Rate ) https://cloud.baidu.com/product/pctr.html
Ad pCTR service is based on Baidu BML. Feature selection is based on rich experience of Baidu Ad system. Itprovides flexible configurations. For example, customers can select features set freely and it also providesmany known effective and efficient algorithms, such as distributed Logistic Regression, Factorization machine,etc.
Baidu Recommender System https://cloud.baidu.com/solution/recommender.html
Recommender systems have become increasingly popular in recent years, and are utilized in a variety ofareas including movies, music, news, books and products in general. Recommender systems are a usefulalternative to search algorithms since they help users discover items they might not have found otherwise.Baidu Recommender System is an end-to-end solution for our business customers to recommend news,movies, etc to their own personal customers. It provides many known effective and efficient algorithms, suchas User/Content Based Collaborative Filtering, distributed Logistic Regression, Wide and Deep, etc.
Baidu ABC Appliance http://tech.sina.com.cn/roll/2017-09-17/doc-ifykyfwq7522033.shtml
Baidu ABC Appliance is a cluster of servers with GPU cards. Each server typically has 8 ~ 16 Nvidia GPUCards, say P40. We developed a distributed software stack named APE. APE runs on the cluster and providesmulti-tenant access. Each tenant could be assigned different quota, e.g. a tenant can only use 10 GPU cardsat the same time. A tenant could run multiply deep learning training jobs at the same time only if his quota isnot used up. In China market, we already have dozens of customers, most of them are leading enterprise intheir industries, e.g. China Life Insurance, Baosteel.We also design and develop deep learning model for our customers on ABC Appliance, e.g Deep learningbased Computer Vision model for Steel Sheets Quality Inspection.
Baidu BDL (Deep learning platform) https://cloud.baidu.com/product/bdl.html
Baidu MapReduce provides customers managed clusters to run deep learning training jobs on Tensorflow,PaddlePaddle or other deep learning frameworks. The virtual machines in the cluster can be configured withdifferent number of GPUs to accelerate the training jobs. BDL is the first in products of this type in Chinamarket.
Baidu Message System https://cloud.baidu.com/product/kafka.html
It provides a unified, high-throughput, low-latency, multi-tenant service on Baidu cloud for handling real-timemessages. Its transport protocol is 100% compatible with Kafka. It is easy for many customers who are familiar withKafka to transfer their projects to Baidu cloud. It now proceeds more than 200 billions messages per day and neverlose any message.
Baidu ABC Robot OSRobot for commercial purpose is a new focus area for Baidu. The goal of ABC Robot OS is to provide manyrobot manufacturers an out-of-box AI platform. ABC Robot OS integrates Baidu AI techniques including ASR(Audio Speech Recognition), TTS (Text To Speech), Face/Object recognition, Indoor location/navigation, etcinto one platform. It is much easier for robot manufacturers to develop a new model with strong AI capabilities.Our customers includes leading enterprise such as UBTECH.
Baidu MapReduce https://cloud.baidu.com/product/bmr.html
Baidu MapReduce provides customers managed clusters to run Hadoop/Spark optimized by Baidu. Customerscould easily do big data analysis and mining on Baidu Cloud.
Baidu BML (Machine learning as a service) https://cloud.baidu.com/product/bml.html
BML provides customers machine learning as a service. The machine learning algorithms are distributed anddeveloped/optimized by Baidu. Customers could use this service to do feature engineering, data transforming,model training, model serving, etc. It aims to cover full life cycle of machine learning project.
2010.12 ~ 2015.12: Senior Software Development Engineer at GoogleKnowledge Graph
Google Knowledge Graph (KG) project is huge and beautiful effort on modeling general knowledge. Unlikesearch engine that organize knowledge based on web pages/documents, KG models truth knowledge byentities and relationships between entities. It is easy to do knowledge retrieval and inference on KG. It is animportant engine of Google's search system and responsible for a significant portion of all queries. I was techlead of Beijing team of KG Collections team and was responsible for learning to define entity collections andlearning to rank members in KG collections.
Books Search and Books Vertical in Knowledge Graph (KG)Books Search (books.google.com) is the biggest book search engine in the world. Books Vertical in KG is oneof the most important verticals. Books metadata (such as book title, book authors, book publishers, bookabstract, etc) is the key in the above 2 projects. I was tech lead of Beijing team and responsible for building thepipeline to process book metadata, convert book metadata into triples and export them to KG. The pipeline isvery long, as it must handle crawling metadata from more than 100 partners/sources, converting metadata ofdifferent format to standard ProtoBuffers, clustering the same book from different partners/sources into one,generating triples, exporting to KG. I also developed an online service which takes as input a big chunk of text(such as book content) and within 1~2 seconds retrieves all books with similar contents from more than 100millions books. It is more challenging to some extent than looking up similar web documents, as book contentis typically much longer than web documents.
Indoor LocationIt was a Google X project and now a feature of Google maps. It is a machine learning project. We collectedWIFI fingerprints by site-survey, then we built a regression model to predict Geolocation (latitude, longitude) fornewcomers with phone indoor. At that time, the model accuracy was about 4 meters and was launched inGoogle maps.
2008.7 ~ 2010.12: Software Development Engineer at Microsoft MBDC
I was responsible developing Office Lync 2010 management server.
Presentation and Cooperation
"Create 2017" Baidu AI Developer Conference, I gave a presentation << 5 Platforms for AI Developers >>
The 5th Baidu Tech Open Day, I gave a presentation << Baidu Cloud Deep Learning Platform
ISC 2017 at Frankfurt Germany ，I was the judge and the author of AI Competition
- Familiar with deep learning theory and well known deep learning structures, such as CNN, RNN, GAN, etc
- Familiar with Google distributed computing/storage tools/projects: Bigtable, MapReduce, Dremel, etc
- Familiar with open source distributed computing/storage tools/projects: Hadoop, Spark, Spark, Kafka, etc
- Familiar with machine learning theory and well known algorithms✧ Familiar with Java, C++, Python, Go
- Familiar with Docker and K8S
- Fast learning on newly advanced technologies. Adaptive to various stressful environments.
- Hard-working and dedicated. Good interpersonal skills and teamwork orientation.
- Good at starting a project/team from 0
- USA Patent: Generating an indoor map model (Google Patent Number: US9052206) http://t.cn/RLh6DpQ
- USA Patent: Navigating Using an Indoor Map Representation (Google Patent Number: US20140113665)
- USA Patent: DETERMINING COLLECTION MEMBERSHIP IN A DATA GRAPH (US20150100605)
- USA Patent: AUTOMATIC DEFINITION OF ENTITY COLLECTIONS (US20150100568)