Speaking

A deep technical walkthrough of tensor parallelism, pipeline parallelism, and sequence parallelism — what they are, when to use each, and the real tradeoffs at scale. Includes live architecture comparisons.

Why KV cache management is the most underappreciated problem in LLM serving, how PagedAttention solved it, and what the next generation of solutions look like.

Moving from demos to production agentic systems is a different engineering problem than people expect. A practitioner's guide to the failure modes, architectural patterns, and operational challenges of real agentic AI.

How to design ML platforms that actually serve the organization — from feature stores to model registries to inference infrastructure. What works at scale, what doesn't, and why.

The most important architectural decisions you'll make when taking a GenAI application from prototype to production — and how to make them well. Covers model selection, serving strategy, evaluation, and ops.

A data-driven talk on the actual cost structure of LLM serving — compute, memory, network, and everything else. How to model costs, where the leverage is, and how to build cost-efficient LLM products.