Tejas Bhakta is founder and CEO of Morph, which builds models and infrastructure for coding agents. His technical work focuses on making inference and the supporting operations around an agent faster and more reliable.[1][2]
Start with Autoresearch Made Our Models 3x Faster — Tejas Bhakta, Morph →
X Career
Bhakta describes himself as a hardware tinkerer and software developer. His company, Morph, concentrates on the operations that support coding agents, including applying edits and finding relevant code. The separation gives a specialized model a narrower job while the main agent decides what work to do. His personal blog connects that company work with a longer interest in building software and hardware.[1][3]
Contributions
In his autoresearch talk, Bhakta describes a repeated loop: an agent proposes a kernel change, a harness checks correctness and measures speed, and the system keeps or reverts the result. Humans supply the larger diagnosis, such as excessive data movement; agents explore implementation parameters. The harness must describe both the actual GPU and the model operation, so the search optimizes a usable implementation.[2]
Ideas
Bhakta warns that improving an isolated kernel benchmark can make overall inference worse, for example by disabling CUDA graphs. His reported threefold improvement combines kernel and hardware changes; the talk does not supply enough benchmark conditions to predict the result for another workload.[2]
Tools, writing & talks
Writing, research & talks
Start here
Explains agent-driven kernel tuning and why correctness and end-to-end measurements both matter.
Watch this if you understand GPU inference and want to automate optimization experiments. Focus on the keep-or-revert tests and the warning about local benchmarks before trying the loop on your own kernels.
Speaker[2]Original source