Design and run experiments to measure how architecture decisions affect inference; create architecture variants optimized for inference speed; own post-training pipelines (fine-tuning, evaluation, adaptation); scale inference for large MoE models; publish research and build agent-driven research tooling.

Design, implement, and optimize extremely low-latency GPU kernels and a monokernel inference pipeline for LLMs across AMD and NVIDIA. Perform microarchitectural experiments, profile and instrument kernels, build in-kernel profiling, scale to large models (MoE), and develop autonomous AI agents for GPU engineering research.