A photograph contains a surprising amount of safety information. It also leaves a lot out. I built HazardAI to see whether a vision model could turn job-site photos into a useful first draft of a job hazard analysis without giving the model authority it should not have.
The app is a mobile prototype for natural gas utility work. A field worker captures one or more photos, and the model returns a structured description of the scene, potential hazards, severity and confidence ratings, recommended mitigations, and conditions that appear compliant.
A draft, not a decision
A vision model can notice missing protective equipment, equipment near an excavation, or poor housekeeping. It cannot reliably measure trench depth from a photograph, know what happened outside the frame, or confirm that a required procedure took place. A polished response does not change those limits.
That shaped the workflow. HazardAI asks a person to review every finding. The assessor can confirm it, edit it, dismiss it with a reason, or add something the model missed. The assessment cannot be submitted until that review is complete and signed.
Keep the disagreement
I wanted the original AI result to remain visible after a person made changes. The prototype keeps the model's analysis separate from the reviewed version, along with the reasons for edits and dismissals. That creates a useful record of where the model helped and where human judgment corrected it.
This matters for more than auditing. Those disagreements are the beginnings of an evaluation set. Over time, they could show which kinds of hazards the model recognizes consistently and where it creates false confidence.
The field changes the product
Safety work does not always happen with a reliable connection or in comfortable conditions. HazardAI stores photos and assessments locally, queues changes for later sync, and uses large controls and high-contrast colors intended for field use.
Those choices are less flashy than the vision model, but they determine whether the idea could be useful outside a demo.
What comes next
The current prototype can call Gemini, OpenAI, or Azure OpenAI directly. It asks the model for policy references, but those references are not yet grounded in an approved policy store. The next meaningful step is retrieval against company and regulatory documents, followed by testing on a reviewed set of job-site images.
I started with a simple question: can vision AI help someone notice a hazard they might otherwise miss? Building the prototype clarified the boundary. The model can prepare a first pass. A person still owns the judgment.
Join the conversation
Thoughtful responses are welcome. Comments are reviewed before they appear.