• Built an AI agent (Python, Gemini, Next.js) for Suspicious Actor Report generation. Orchestrated agentic workflows for risk signal identification, data aggregation, and report drafting. Built a copilot interface that allows investigators to review and correct AI-generated analyses at each stage. Improved report accuracy by 50% through prompt engineering and workflow optimization.

• Built an end-to-end PySpark pipeline for extracting, summarizing, and clustering risk factors from risk reports using Gemini. Productionized the pipeline from research notebooks onto PayPal’s ML platform and achieved 10x reduction in run time through batched LLM calls and parallel execution.

• Evaluated and integrated vLLM and LMCache into production ML pipelines, reducing inference latency by 40%.

• Built and maintained 3 core infrastructure components powering PayPal’s rewards platform (300M+ users): reward program management, real-time transaction processing, and batch payouts.

• Rebuilt the team’s test infrastructure by enabling parallel execution and migrating test cases from Excel to Java. Reduced test execution time by 18x and improved traceability of test case changes.

• Improved production stability for the rewards platform’s distributed services by triaging cross-team issues, debugging data pipeline failures, and resolving production incidents.

• Decoupled code and content for PayPal Shopping’s codebase and integrated WorldReady Service (PayPal’s localization package) to store and serve content remotely; enabled live update for content change and reduced release latency by 50%.

• Created command-line tool that generates consistent JSON payload for piping through hash functions; published the tool as open-source npm module and received 100+ downloads in the first week.

• Used Docker to containerize TensorFlow models powered by WasmEdge, Second State’s WebAssembly virtual machine, and deployed service on AWS Lambda and Tencent SCF; wrote detailed tutorial on CNCF’s blog.

• Developed chatbot prototypes with Rust and deployed on WasmEdge-powered FaaS platform to showcase its speed and convenience; created user documentation on usage of the plaform with code examples.

• Researched on integration of WasmEdge in smart cars (with tools from the Erdos Project) and development of WebAssembly containers (with CRI-O).

• Collaborated with engineering team on designing credit system for TikTok advertisers; used Pandas and Matplotlib to validate the system and visualize credit score distribution.

• Processed and analyzed behavioral data of advertisers with SQL and Python to detect traits that mark off risky advertisers; implemented corresponding risk-control features and reduced key risk metrics by 60%.

• Built shared monitoring dashboards for key risk metrics to accelerate information transmission between engineering, operations, and sales teams.