This talk explores building a complete self-hosted LLM stack in Rust: Paddler, a distributed load balancer for serving LLMs at scale, and Poet, a static site generator that consumes those LLMs for AI-powered content features.

This talk explores building a complete self-hosted LLM stack in Rust: Paddler, a distributed load balancer for serving LLMs at scale, and Poet, a static site generator that consumes those LLMs for AI-powered content features.
We'll dive into the hard problems: async request routing across dynamic agent fleets, integrating with llama.cpp's C++ codebase, managing KV cache in custom slots, and implementing zero-to-N autoscaling with request buffering. You'll see how Rust's ownership model prevented entire classes of bugs in distributed state management, and walk away with concrete patterns for building and consuming LLM infrastructure in production.
During this talk we'll build a basic, working async runtime using nothing more than a standard library. The point? To see it's approachable for mere mortals.
In my session, I will present the https://hotpath.rs crate and explain how it compares to other profiling tools available.