I work at DoorDash Labs, the robotic and automation arm of DoorDash. Currently focusing on perception and prediction.

Led the object detection team in perception, which consisted of 10+ engineers and two line managers.

The team is responsible for the on-vehicle 3D object detection system, machine learning model training, optimization and deployment, as well as data preprocessing and model evaluation.

Main focus areas include LiDAR-based and camera-based detection, sensor fusion, latency optimization and model training automation. My team also worked closely with partner teams to resolve common on-road perception disengagements and improve the system-level driving performance.

Tech lead manager in Perception-Prediction at Uber ATG. Led a team of 5 based in the SF office. Developed, deployed and improved key perception and prediction deep learning based models on ATG's autonomous vehicles.

Major past projects include

• Deep learning based motion forecasting for road users (vehicles, pedestrians and cyclists) around autonomous vehicles.

• End-to-end trained joint perception-prediction deep model that generates object detections and motion predictions from raw sensor inputs.

• Sensor fusion (Lidar, camera, radar) in perception-prediction models.

The above projects also led to multiple patents and conference publications (ICRA, IV, WACV, KDD, etc.)

• As a founding member of the revenue optimization team, helped it grow to a 10-person team tackling a wide variety of applied machine-learning (ML) projects.

• Built a ML based sales recommendation system, which increased sales productivity by 10%.

• Built an accurate ML model to predict advertiser churn (84% ROC-AUC).

• Led a project to optimize mobile site content with ML, which increased app installs by 7 %.

• Led the design and implementation of a feature data store and ETL pipeline shared by multiple ML systems.

• Helped organizing a Kaggle recruiting contest on Yelp photo classification; served as an interviewer and a mentor; presented in engineering learning groups and internal ML tech talks.

Developed a bottom-up simulator to model the real-time bidding of online ad, by analyzing the massive bidding data. During this project, I developed a modified logistic regression model that can handle periodic input features (e.g. time of day). The simulator was used to quickly evaluate new bidding strategies before A/B tests, and to project the budget spending accurately. The click-through rate model I revised led to a 10% revenue lift during an A/B test that I co-run.