Co-founder of MLPerf, the industry standard benchmark for ML systems.MLCommons is an open engineering consortium dedicated to improving machine learning for everyone. We have launched open datasets and tools to drive democratization of machine learning https://mlcommons.org/en/news/neurips21/

Founded an LLM systems startup. We invented a new finetuning technology to remove LLM hallucinations. We deployed to more than 10 enterprise contracts, 50 design partners, and 15,000 developers. Raised $26M and led a team of 30.

Figured out how to build LLM systems on AMD HW (MI100, MI200, MI300), prototyped with emerging AI startups.

I learned how to create and work with startups by working with AI Fund companies, including Landing AI and Netail.

Built and programmed the first 1000 CUDA GPU cluster to train a deep learning model that was shipped at scale to 100s of millions of users.Discovered scaling laws, motivating large language models - shipped them in the Baidu search engine: https://arxiv.org/pdf/1712.00409.pdfFounding member of the Silicon Valley AI Lab, contributed to the DeepSpeech and DeepVoice systems. Lead a team of 40 systems and machine learning researchers. Deep Speech was named a top 10 breakthrough technology of the year by MIT Tech Review. We achieved state-of-the-art results in noisy (low SNR) English speech and challenging English conversational speech using deep learning and a high performance training system.

Explored CUDA for many different applications, including deep learning.

Invented the independent thread scheduling system that shipped in the Volta GPU.

Wrote the compiler for DARPA UHCP project Echelon, based on Ocelot.

Wrote an Ocelot backend for the GEN GPU.

Wrote an Ocelot based code generator for a GPU based SQL database engine. My first startup experience.

Interned in John Nickoll's CUDA architecture group. Wrote the first version of the Ocelot compiler.

How I learned to stop worrying and love the machine.