Building Production Data Agents. Design, Evaluate, and Scale AI Workflows for Data Analytics (en Inglés) - YI AI

A production data analytics system built with language models requires more than prompt-based SQL generation.
A usable system has to do more than translate a question into a query. It needs to represent schema meaning,
retrieve the right context, handle ambiguity, validate queries before execution, protect sensitive data,
observe failures, evaluate changes, and improve over time. In practice, those concerns determine whether a
workflow is reliable, safe, and maintainable.

This book walks through the design and implementation of a production-oriented SQL agent workflow from the
ground up. The focus is not a toy demo or a single prompt. It is the full system around the model: schema
representation, retrieval, structured generation, agent loops, security controls, observability, evaluation,
and model optimization for real production domains.

You will learn how to build workflows that can interpret business terms, retrieve the right context for each
request, generate and validate SQL safely, recover from failures, manage memory across interactions, and expose the signals needed for monitoring and continuous improvement. The later chapters cover prompt optimization,
fine-tuning, reinforcement learning, and workflow-level design choices that improve the behavior of the whole
system.

The implementation uses a practical open-source stack built around Python, LangGraph, LangChain, Hugging Face models, Unsloth for efficient fine-tuning, and VERL with Agent Lightning for reinforcement learning workflows. It also integrates Langfuse for observability and tracing, enabling tracking of agent execution, prompt interactions, model outputs, and system performance during development and evaluation.

This book is for developers, data engineers, and ML engineers who already know basic SQL and want to
build production SQL agent systems from scratch.