Tejas Bhakta

Tejas Bhakta was introduced in his 2026 AI Engineer World’s Fair biography as the founder and CEO of Morph, building specialized models and specialized inference for code generation. He previously worked as a machine learning engineer at Tesla, where he worked on inference optimization for Tesla AI.

In his presentation, Bhakta describes how his team uses autoresearch to optimize custom GPU kernels, including for lower-cost GPUs without NVLink. Humans identify optimization opportunities, while agents tune block sizes and other parameters through repeated correctness checks, performance benchmarks, and decisions to keep or revert changes. His approach gives agents context about the specific hardware and model architecture and evaluates end-to-end inference performance to catch misleading improvements, such as a faster isolated kernel achieved by disabling CUDA graphs. He reports a combined threefold speedup from kernel and hardware optimizations, with bare-metal tuning contributing roughly 25% over a virtualized setup, while emphasizing that custom kernels have workload-specific limits and roughly 80% of autoresearch attempts are unsuccessful.