Architected a 20B-parameter Vision-Language-Action model trained on 8K+ TPUs, the first foundation-model-based autonomous driving stack. Designed the multimodal interface connecting Lidar and Camera encoders to reactive and deliberative reasoning decoders.

Led Waymo's first TPU inference deployment, growing to 15+ models company-wide. ~6x latency reduction for VLMs (145ms → 25ms), >50% HBM savings, >$150M annual cost reduction. Custom kernels in JAX/XLA/Pallas.

CNN-based framework for identifying solar installations from satellite imagery. Used to construct the first detailed map of solar adoption in the United States.