Project collaborators
Mitch Swartz, SafeSet
Maria Oros, Data Science Institute
Project start and end dates
March 2025-June 2025
Project summary
“From the start of our AI project, it felt like working with a true thought partner. The team took the time to understand both the technical side and the safety context we were building for, and they were open to iteration as the idea evolved. I’d highly recommend pointing other founders your way, especially those working on AI applications aimed at complex, real-world problems.”
—Mitch Swartz, SafeSet
Manufacturing reliability and safety are compromised by human error arising from workforce factors like fatigue, cognitive overload, and inconsistent scheduling, leading to production failures and safety incidents. In this industry collaboration with SafeSet, DSI developed a novel error trap detector utilizing the latest advancements in large language models (LLMs). This system is designed to autonomously analyze time-series workforce scheduling data to proactively identify anomalous patterns associated with elevated risk.
DSI led the data analysis, infrastructure design, and model development, including the deployment of an in-house LLM on Databricks for scalable inference and experimentation.
The system integrates in-context learning and prompt-based reasoning to detect operational anomalies and explain underlying risk patterns. A comprehensive benchmark evaluation assessed model performance against conventional machine learning and rule-based approaches. In parallel, a literature review and experimental research study identified optimal strategies for LLM-based arbitrage, such as majority voting and ensemble reasoning, to enhance prediction robustness.
This effort translates complex predictive analytics into actionable, human-centered alerts for safety monitoring. The initiative drives social good by creating a safer, more efficient manufacturing environment through a data-driven, early-warning approach.
Project deliverables
- LLM power-based application for anomaly detection
- Statistical analysis