• Integrated DeepSpeed to Bpex’s benchmarking tool to use distributed training and ZeRO to support larger models.
• Wrote high-performance CUDA kernels for GEMM that utilize OP fusion, memory hierarchy and tensor cores.
• Profiled performances of various attention implementations, analyzed in depth the differences between xFormers’s cutlass implementation and flash-attention(v1, v2), and delivered a comprehensive report in an internal presentation.

• Developed and maintained a Teams Bot utilizing Microsoft Bot Framework and .NET core, which enables users to effortlessly query databases and request internal APIs using natural language.
• Implemented QuickAccess, an admin web application to easily manipulate and search tax records using ASP.NET core MVC and Azure Cosmos DB.
• Built tools utilizing Azure DevOps SDK to automate routine jobs including a daemon synchronizing data between task items and pull requests and a crawler maintaining data source for the Teams Bot, saving 10 working hours/month.

Participated in developing RisingWave, the next-generation cloud native streaming database.

• Refined system’s value encoding.
• Implemented common table expressions for front end.
• Implemented internal table catalog, including inference and storage.

• Adviser: Prof. Cheng Tan
• Built a benchmarking tool to measure different machine learning models’ performance on Multi-Instance GPU.
• Implemented a “exchange-and-compact” development transition algorithm on top of Kubernetes.
• Optimized the transition time from 30’ to 1'.
• Built the testbed containing 4-instances on GCP and carried out a 24-hour end to end test.

• Adviser: Prof. Yubin Xia and Ph.D. Mingyu Li
• Implemented a secure personal data analysis and storage system on Hikey960 using OPTEE and AOSP.
• Designed and developed an app to generate SJTU students’ personal annual report.
• Participated in Shanghai “Internet+” College Student Innovation and Entrepreneurship Contest as an SJTU representative.