Yuntao Peng

Beihang University (BUAA), Beijing, China

Software Development Engineer

Our team has two scopes. The first is JP prime, in which we cooperates with bussiness and PM teams to develop Amazon prime service for Japan localized needs. The other is customer engagement. We build a platform to power personalized customer engagement experience. We recommend personalized actions for customers and track their behavior to reward them for deeper engagement. I participated in both parts, developing and maintaining new features and doing tech designs in serverless fashion.

Customer Engagement Platform

• [Do X, get Y Project] Developed and launched a new customer engagement project, including landing page implementation on Amazon main site, management web service for project managers to manage offers, customer action tracking system (serverless), contract fulfillment and rewarding workflow, and a aggregator service exposing APIs for operation
• [AWS migration] Designed and developed a serverless based data processing stream to ingest all kinds of customer action, as a part of migration from server-based services to native AWS

Back-End Software Engineer

Search Architecture Department is responsible for maintaining Baidu search offline web page library, web index library and online real-time search service. My team is in charge of the entire back-end architecture. I mainly participated in the build of back-end intelligent architecture. We developed and maintained an online model inference framework. Through the mixed deployment of models, one instance can respond to multiple inference requests, which greatly improved the utilization rate of computing resources; through platform-based design, the deployment details were shielded from access services, so that only simple RPC requests are required to obtain highly stable model inference services with success rate of 99.9%.

Model Service: online heterogeneous computing resource (CPU, GPU, XPU) inference server

• [Dynamic Retry & Automatic Degradation] Abstracted the proxy layer of the service, and implemented a dynamic retry mechanism based on waiting time per request in the computing service. This mechanism reduced total hardware utilization by 20%. Based on it, implemented a heartbeat detection service to generate degradation parameters according to the response time and service accessibility and send it to upstream applications to realize an automatic delay-based degradation mechanism
• [Model Optimization & Access Support] Optimized Transformer based models (Bert, Ernie) according to hardware types (Tesla T4, Tesla P4, Kunlun K200) using TensorRT and got 50% speedup in average; supported different types of models to access Model Service; provided daily technology support and abnormality response

Based on AES (Advanced Encryption Standard) and SM4 (Commercial Cryptography Standard), I designed two novel parallel encryption algorithms and implemented them on the Nvidia GPU with special design of data transmission optimization

• Designed two parallel encryption algorithms based on Electronic Codebook (ECB) mode, and implemented them on Nvidia GTX 965M GPU.
• Designed experiments to verify the advantages of the parallel encryption algorithms in terms of running time, speedup and throughput compared with the baseline strategy (serial encryption algorithm), and got great improvement shown as below

Repairing Human Trust with An Attention Transfer Model in Robot

To maintain human trust, in this research, a trust repair framework was developed based on a human-to-robot attention transfer H2R-AT model and a user trust study

• Conducted research on the human trust rebuilding mechanism in HRI (human-robot interaction)
• Improved the NLP algorithm with LSTM (long short-term memory) network in the model H2R-AT
• Co-first author of the paper sumbitted to Autonomous Robots (under review)

Human-to-Robot Attention Transfer for Robot Execution Failure Avoidance Using Stacked Neural Networks

It is challenging for a robot to detect its failures or failure tendency by themselves. In this research, we focus on using human attention to remind robots of its potential failures, in cooperation such as daily assistance and cooperated manufacturing. A novel deep learning architecture will be designed to accurately understand human instructions, and then assist with robot behavior corrections.

• Proposed a new method, H2R-AT(human-to-robot attention transfer) to identify and correct robot abnormal executions by sharing human attention based on stacked neural networks
• Designed the simulation platform for the project on simulation software V-REP using Jaco Arm and OMPL(open motion planning library)
• Designed the experiments and user study. Collected and classified data from online questionnaire
• Co-first author of the paper published on IEEE Transactions on Human-Machine Systems