| Abstract Scope |
Manufacturing process data are rarely analysis-ready: they contain missing values, label noise, recipe changes, lot histories, equipment drift, and undocumented shop-floor context. This work proposes a domain-grounded agentic AI framework that autonomously orchestrates the end-to-end workflow from data ingestion and preprocessing to task identification, feature engineering, model selection, hyperparameter optimization, training, evaluation, and quality/defect prediction. The framework combines tool-using LLM agents with Python-based analytics, machine-learning libraries, visualization tools, and retrieval-augmented domain knowledge from process standards, equipment records, and expert rules. To avoid treating autonomy as an unchecked black box, the workflow records tool calls, data transformations, assumptions, model decisions, and evaluation claims for human review. A manufacturing quality-prediction case study is used to compare the agentic workflow with a human-defined baseline in terms of preprocessing consistency, predictive performance, error detection, reproducibility, and traceability. The results suggest a practical path toward reusable, auditable AI workflows for manufacturing data science. |