Why does nobody talk about AI training data for niche trades?

Last month I was chatting with a welder in Tulsa who wants to build a little QA tool for his shop. He asked me if there's any way to get AI to spot porosity in welds from photos. I looked around and everything I found is either generic object detection or some fancy industrial system that costs 20 grand. There's no middle ground for a guy with a phone and 500 sample photos. Three years ago I tried something similar for my own garage cabinets, giving ChatGPT a bunch of my drawings and asking it to flag mistakes. It got the obvious stuff but missed the exact things I cared about, like hinge alignment or plywood grain direction. Seems like the real gap is getting the AI to learn from small, messy, real-world datasets instead of huge clean ones. Has anyone actually gotten good results training a custom model on their own trade-specific photos without a data science degree? And if so, what did you use?

1 comments

1 Comment

ruby_bell472d ago

You're talking about a welder in Tulsa with 500 photos and you think AI is gonna crack that? I get the appeal but it feels like people overthink this. A dude with a phone and some samples is better off eyeballing the welds than fighting with a model that probably spits out nonsense half the time. The real issue isn't the tools, it's that these projects sound cool but they rarely save anyone actual time. I tried custom training on my own stuff once. Gave up after a weekend. The juice wasn't worth the squeeze.