I build self-improving AI agent systems at NVIDIA, and I have worked on AI acceleration and infrastructure security.
My task agents implement, critique and verify, while an overseer meta-agent watches for drift and weak tests.
At CentML we built a fully fused GPU kernel for an entire MoE layer, speeding computations up 2 to 11x.
One agent implements, the other critiques and verifies against my PRD. It loops until both agree it's done.
Over an 8-hour agentic run at home I measured a 95.8% prefix-cache hit ratio. In AI, cache is king.
I hosted a meetup and open discussion on vLLM, the inference engine, on a Thursday in September.
I build self-improving AI agent systems at NVIDIA, and I have worked on AI acceleration and infrastructure security.
My task agents implement, critique and verify, while an overseer meta-agent watches for drift and weak tests.
At CentML we built a fully fused GPU kernel for an entire MoE layer, speeding computations up 2 to 11x.
One agent implements, the other critiques and verifies against my PRD. It loops until both agree it's done.
Over an 8-hour agentic run at home I measured a 95.8% prefix-cache hit ratio. In AI, cache is king.
I hosted a meetup and open discussion on vLLM, the inference engine, on a Thursday in September.
By using this service, you agree to the Terms of Service and Privacy Policy.