Reinforcement learning for LLMs still looks like something only big labs can afford: expensive GPUs, complex training stacks, fragile recipes, and weeks of trial and error. But the ecosystem has quietly changed. With LoRA-based training, TRL, Unsloth, and adapter-aware serving in systems like vLLM and llama.cpp, it is now realistic for ordinary engineers to run useful GRPO experiments and serve multiple task-specific adapters on shared infrastructure.

This talk is for engineers who want to move beyond prompting and try practical RL post-training without building a research lab. We’ll walk through a small GRPO training recipe using LoRA, explain the 5–6 hyperparameters that matter most, and show how those choices affect training behavior in practice.

Attendees will leave with a runnable notebook, a mental model for configuring GRPO runs, and a practical path from “trained LoRA adapter” to “servable model variant” on existing inference infrastructure.

Prompt optimization is still mostly hand-crafted: edit the wording, rerun a few examples, inspect failures, and repeat. That is like hand-crafting image kernels for every new vision problem. It can work for simple cases, but it does not scale well, and it is hard to reproduce.

GEPA offers an algorithmic alternative. It is a gradient-free form of RL for LLM systems: sample candidate behaviors, evaluate them, use feedback from failures, and search for better policies. But instead of updating model weights with PPO or GRPO, GEPA optimizes the text artifacts that shape model behavior: prompts, and system instructions.

This talk is for engineers building LLM applications or agent workflows who want to move from prompt tweaking to evaluation-driven optimization. We’ll walk through the core GEPA recipe: writing an evaluation callable, returning both reward and natural-language feedback, exposing one or more text variables to optimize, and running the optimizer on a small dataset.

Attendees will leave with a runnable notebook, a mental model for when GEPA is useful, and a reusable recipe for applying automatic prompt optimization to their own LLM systems.