Get the eBook free when you register your print book at Manning. It's a fundamental truth that all software--even AI systems--is broken. AI engineers who can diagnose faults and refine systems to align with business needs are in high demand. This book expands the foundational research into judging and adapting AI systems into a collection of practical techniques you can use on the job. As you trace the progression from surface-level text matching to semantic similarity to judgment-based evaluation, you'll build the mental models necessary to choose the right metrics, detect failure modes, and close the loop from evaluation to alignment.
This book teaches you to think of evaluation as a design constraint. You'll employ a "working backwards" methodology that begins with what your system must get right, which directs you to the appropriate evaluation approach. As you internalize the define > evaluate > analysis > align cycle, you'll start making more informed tradeoffs and expertly balancing helpfulness, safety, and brand voice in your models.
What's inside - BLEU, ROUGE, BERTScore, COMET, and LLM-as-a-judge methods
- Detecting and quantifying hallucinations
- Aligning AI with RLHF, constitutional AI, and red teaming
- Timeless best practices that will apply as models evolve
About the reader For AI engineers and LLM practitioners. No prior knowledge of NLP metrics, reinforcement learning, or alignment research is required.
About the author Han Lee has spent more than a decade applying cutting-edge research on large-scale AI and machine learning systems into production-grade products. A Senior Director of Data and AI at Moody's, he leads teams that ship generative AI applications and has daily, hands-on exposure to safety-critical evaluation pipelines.
