stanfordnlp/dspy ? reverse-engineered prompt

I want to build a new Python framework for working with large language models. The main idea is to move away from "prompt engineering" and towards a more structured, programmatic approach.

Instead of writing brittle, complex prompts, developers should be able to define the steps of their AI system, like a RAG pipeline or a simple agent, using simple, modular Python code. The framework's job would be to take this high-level program and automatically optimize it. It should be able to figure out the best prompts for each step on its own, and maybe even handle fine-tuning. It's essentially a compiler that turns declarative Python logic into a self-improving LM pipeline that produces high-quality results. Let's call it DSPy.