I like to just do things.

Trained models to generate floor plans, designed bespoke geometry-processing algorithms, and built infrastructure capable of supporting many concurrent users.

Worked with NVIDIA to optimize DeepSeek R1 inference, setting the world performance record at the time and outperforming specialized hardware systems.

Built custom expert-selection and routing kernels, custom GEMM kernels, and a new quantization specification engineered to minimize error.

Trained transformer models across unusual modalities—building systems beyond the familiar boundaries of text and image.

Built the first platform for fine-tuning and deploying models such as GPT-J, pioneering multi-LoRA serving roughly a year before open-source implementations caught up.

Reached $131k monthly peak revenue and built a chat product—with early innovations including chat sharing—that grew to 3 million users.

Developed state-of-the-art wildfire spread prediction models using ViT U-Nets for California and Colorado state wildfire services.

The earliest layer: building a company from zero and establishing the product instincts underneath every system that followed.