• Implemented custom CUDA kernels for perception models and preprocessing stages.

• Model quantization with mixed layer precisions to balance latency and quality.

• Model deployment on edge device using TensorRT.

• Various kernel level and scheduling optimizations to maximize GPU usage & reduce inferencing latency.

• Remodeled our distributed cluster coordinator to enable fast failover in seconds at million object scale, implemented distributed transitions to resolve consistency problems.

• Scaled asynchronous replication objects by 10x while preserving original RPO goal.

• Worked on the data path for near-sync replication to enable sub-second disaster recovery.

• Various performance optimizations involving lock optimization, spurious wakeup, cache alignment etc, to ensure performance at scale.