• Designed and implemented a custom ticket(a collection of workflows for algorithm estimation) template feature, allowing users to conveniently reuse it and reducing the error rate of original ticket submissions.

• Developed RESTful APIs for CRUD operations on ticket templates, interacting with a PostgreSQL database. Created and implemented unit-tested APIs to launch tickets from customized templates. Deployed these APIs to the cloud using Kubernetes, significantly boosting ticket submission efficiency.

• Migrated hard-coded ticket functions to a database-driven template system, streamlining the process of modifying auto-launched tickets for various algorithms. Additionally, implemented visible and writable authentication and fuzzy search capabilities.

• Connected the learned paging framework to the Structure Simulation Toolkit (SST) as a dynamic shared library. Replaced the radix page table walker with our learned page table in SST, utilizing a multi-level set associative cache to store virtual addresses, thus simulating and validating the algorithm’s effectiveness.

• Designed and implemented a memory allocator using the best-fit policy, enhancing the efficiency of memory allocation and deallocation.

• Developed various scripts to parse and print experimental results concurrently, compress and decompress simulation data, and plot experiment outcomes, enabling efficient analysis and visualization of the learned page algorithm’s effectiveness.

• Designed and implemented the server-to-local application interfaces of the Optical character recognition(OCR) system, boosting the efficiency of data transmission between local app and the server.

• Improved the accuracy of the receipt OCR system in the pre-processing step by denoising the documents.

• Optimized the object detection section of the OCR system by using single shot detector (SSD) model.

• Reproduced and optimized proposed DPML(Differential Privacy Machine Learning) algorithms in a benchmark project.

• Visualized and compared the performance of different DPML mechanisms, making the benchmark a reference for future research.