2025/01 - Present:

• E2E Agentic Tool-use benchmarks for Gemini quality hillclimbing

• Gemini Enterprise agent harness quality research

• Post-training (SFT, RL) on horizontal agent capabilities

2024/02 - 2024/12:

• (20% project) Scalable synthetic STEM QA auto-generation for Gemini 2.0 Pre-training.

2022/07 - 2024/12:

• Ads customer engagement data infra/platforms @ Google Ads

• Joined supply-side platform (SSP) team on Ad Tools platform, enabling ad targeting, ad provider bidding, and ad delivery to games that serve more the 300 million users.

• Boosted resource loading time on Ads Tooling Portal’s listing pages by 500ms through building a server-side preloading pagination component using TypeScript.

• Implemented a generalized Promise loader for asynchronous data handling across all pages on Ads Tooling Portal using React higher-order component.

• Streamlined user segmentation process for ad targeting by building a dynamic visual editor using React with Typescript and MobX state management.

• Created backend methods for uploading user segment configuration JSON to Google Cloud Storage buckets using internal Java framework pFlight.

Co-developed CrimeForecaster, a crime forecasting deep learning framework that captures crime events’ spatial and temporal dependencies via Diffusion Convolution and Gated Recurrent Network, using TensorFlow (Python)

Represented physical and semantic proximities of neighborhoods as weighted edges in graphs through doing geographical computations in spatial databases using PostgreSQL and PostGIS

Improved forecasting performance by 10% via feature selection through testing different feature combinations on Machine Learning baselines using scikit-learn

Coauthored paper “CrimeForecaster: Crime Prediction by Exploiting the Geographical Neighborhoods’ Spatiotemporal Dependencies”, published in ECML-PKDD 2020

Co-developed ReCOVER, an user-interface web application for showcasing predictions from COVID forecasting model SI-kJapha using React (JavaScript) and React Router. Projects got featured by CDC.

Built and maintained the back-end server for collecting and updating worldwide and nationwide COVID infections/deaths cases everyday using Django (Python). Received above 700 requests per day from users.

Co-developed COVID19-Forcast-Bench, a front-end leaderboard app for collecting predictions from 43 different labs and sorting labs on a table based on mean squared error using React

Implemented and beautified front-end forms, graphs and dynamic heat maps using JavaScript libraries including Ant Design, Nivo, amCharts, Recharts, and DevExtreme Reactive

Provided customer service for 5800 number of students and guests through the office’s 11

programs. Facilitated small group discussions for new students regarding academic success, campus resources, university policies, and other common concerns for first-year students.