How generative AI could help make construction sites safer

To combat shortcuts and risk, Lorenzo works on a San Francisco -based Dronzo tool for San Franciscoand Who sells programs that create daily digital models to apply at work from videos and photos, known in trade as “capture”. The tool, called Safety Ai, analyzes realistic images every day to take pictures and flags that violate the OSHA safety and health management rules (OSHA), with what is claimed to be 95 %.

This means that for any danger to safety in the program ads, there are 95 % of the certainty that science is accurate and related to the organization of a specific OSHA. It was launched in October 2024, and it is now published on hundreds of construction sites in the United States, says Lorenzo, and the publications of building systems have been published in countries including Canada, the United Kingdom, South Korea and Australia as well.

Safety Ai is one of the multiple construction safety tools that have appeared in recent years, from Silicon Valley to Hong Kong to Jerusalem. Many of these things depend on the difference of human “vertes”, often in low -wage countries, to draw manually specific squares around pictures of major organisms such as stairs, in order to name large amounts of data to train algorithm.

Lorenzo says that safety of artificial intelligence is the first to use obstetric intelligence for the mark of safety violations, which means an algorithm that can do more than getting to know things such as stairs or solid hats. The program “Think” can about what is going on in a picture of a site and take a conclusion about whether there is a violation of OSHA. This is a more advanced form of analysis than the detection of the object, which is the current standard of industry, as Lorenzo claims. But as 95 % success rate suggests, artificial intelligence is not flawless and knowledgeable. An experienced safety inspector requires an experienced supervisor.

A visual language model in the real world

Robots and AI tend to prosper in fixed, largely fixed environments, such as factory floors or charging stations. But construction sites, by definition, change slightly every day.

Lorenzo believes that it has built a better way to monitor sites, using a type of gym called the visual language model, or VLM. VLM is LLM with a vision encryption, allowing it “Seeing” the world’s “visions” and analyzing what is going on in the scene.

Using years of reality, I collected customer photos, with their explicit permission, the Lorenzo team collected what he calls a “golden data group” that includes tens of thousands of photos from OSHA violations. After he stored this carefully specific data for years, he is not concerned that even the technology giant is billions of dollars, he will be able to “copy and crush it.”

To help training the model, Lorenzo has a smaller team of construction safety positives raises strategic questions from artificial intelligence. The trainers enter test scenes from the Golden Data collection to VLM and ask questions of the form through the process of breaking the scene and analyzing it step by step in the way a person with experience desires. If VLM is not born the correct response – for example, it lacks a false violation or record – the human trainers return and adjust the claims or inputs. Lorenzo says that instead of just learning to get to know things, VLM is taught “how to think in a certain way”, which means that it can extract accurate conclusions about what is happening in a form.