Migration & Modernization
Author: Junjie Tang
Junjie Tang is a Sr. Principal at AWS and Gold Jacket holder, working with the largest automotive and technology companies to build production AI systems at scale. His recent research, NKI-Agent, combines domain-specific fine-tuning with a compile-verify-fix agent loop to generate hardware-optimized kernels for AWS Trainium. His research has been published at ICML 2026 and AAAI 2026. He has spoken at AWS re:Invent every year since 2017 and is an invited speaker at KDD 2026.
Stop hand-porting kernels: migrate PyTorch and Triton kernels to AWS Trainium with an AWS Transform custom agent
An AWS Transform custom agent converts PyTorch and Triton kernels to NKI for AWS Trainium, compiling, verifying, and profiling every candidate on real hardware. The same Claude model that cold-solves 6% of kernels reaches 83% inside the agent’s verify-and-fix loop.