An app that catches safety violations from a single site photo
AI use cases that employees built themselves at a worksite in Korea's largest oil-refining and petrochemical complex were made public. The most striking is a safety-inspection tool: upload a photo of the site and pick the type of work, and it compares the photo against a vast body of work-standard and risk-assessment documents and points out violations. The key point is that it is not a vision model judging whether something is defective but a rule-matching structure in which the AI holds the regulations and people supply only the observations. In the demo it caught inadequate collapse-prevention measures and a missing hard hat, plus a slip hazard from water pooled on the floor, which people had missed by eye. However, the model used, the way documents are referenced, and the accuracy and false-detection rates are not disclosed at all, so it comes with the caveat that it should be read as a catalog for finding projects rather than as an implementation reference.
Key points
- Built because, with 3,000–4,000 people a day entering a complex of 2.5 million pyeong (about 8.3 km²), supervisors cannot go through all the reference documents
- Just uploading a photo and choosing the work type produces output in 2 stages: an analysis against the safety guide, and violations of the risk assessment
- Also presented: a division-of-labor approach that hands off document review first, and running 3 agents at once by loading context in advance and only pressing approve
- At the same event an agent made up nonexistent education and awards, laying bare the risk of unreviewed output
This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.