As semiconductor manufacturing grows increasingly data-driven, fabs are seeking AI solutions that can do more than crunch numbers—they must also interpret complex, unstructured documentation. This talk introduces Agentic AI, a goal-oriented approach that builds on large language models (LLMs) to navigate both structured datasets and the rich, often chaotic world of unstructured text.
Unlike traditional AI pipelines that rely on single prompts or static rules, Agentic AI operates iteratively—asking targeted questions, adapting based on feedback, and collaborating with human experts when needed. This methodology enables AI to work more like a junior engineer: exploring options, learning from context, and solving problems step by step.
We present a case study in which Agentic AI is applied to SECS/GEM equipment manuals—essential documents for tool integration but notoriously difficult to parse. The system dynamically extracts commands, events, and configuration details, significantly accelerating tool onboarding and reducing manual errors.
Beyond this specific use case, the talk explores how Agentic AI can enhance traditional ML/DL pipelines by selecting appropriate models, surfacing relevant features, and even combining symbolic reasoning with neural networks. We conclude with a look at challenges ahead—such as reproducibility, domain adaptation, and human-AI collaboration—and the broader opportunity: embedding reasoning AI across the semiconductor workflow, from commissioning to compliance.
.jpg "Jim Redman photo")