Clear use cases with uncertainties around the solution
A global industrial mining corporation wanted to build applications for internal and external use for two projects with a third potential exploratory project. What typified the use cases were experts looking at images (x-ray and regular) of their products to make business decisions. These experts needed to standardize assessments internally and codify this knowledge to retain and scale their in-house expertise. “PDC cutters are diamond tools used in oil and gas drilling applications. These cutters have essential characteristics such as abrasion resistance and impact toughness that need to be optimized to be effective. These characteristics are measured in lab and production environments by optical and x-ray images.“
The challenge they faced was figuring out how to train a neural network to assess these images, given the specificity of their operating domain. Only a handful of people in the industry understand the context of the images. Although highly technical and experts in their own right, they were not machine learning engineers. Also, they did not want to front-load a significant investment into these projects before understanding the feasibility. They had attempted projects in AI that turned out to be a waste of time and money.
So, how can they efficiently and effectively capture their knowledge in a neural network? How long does it take, and how much does it cost to do so?
Modern tools like Hasty are very accessible for everyone here and allows us to harness the power of AI with a relatively low investment.
Applications engineer from the internal innovation hub
Using machine learning for the heavy lifting of training AI
The labor-intensive part is creating the labeled data asset for custom use-cases. With the machine learning-supported tools in Hasty, it was economically viable to leverage the existing image data asset and the domain knowledge of the subject matter experts to create a substantial ground truth training dataset.
“The Instance segmentation assistant is absolutely fantastic and allows me to use the AI and train it simultaneously. DEXTR and ATOM were my favorite features for labeling; they made the process really fast and accurate. I also really like the keyboard shortcuts and how much thought has gone into putting this together.”
The world-leading diamond maker built the first version of their model with an application engineer, two interns, and a subject matter expert in a couple of weeks. They iterated on this, tracking the performance improvements of the model every step of the way. They could see on the first day that the model was already learning to understand the vital information in the images. The model-assisted labeling was making decent early predictions that improved quickly. Their idea was feasible, and all the team needed to do was put in a few more hours to create enough training data to make the solution reliable and robust. “Modern tools like Hasty are very accessible for everyone here to harness the power of AI with a relatively low investment.”
Having the confidence and internal backing to tackle future cases
“The process is reliably repeatable for a range of use cases where visual information is being used to make decisions.” The management team of the company and key stakeholders plan out future use-cases where the team can use the established process to save the group more money.
“Hasty was the right tool for us, as they automate large portions of building a data-asset which removes large portions of the manual work, improving the ROI and business case. You have built an exciting product that could be useful in many industries, and it could be helpful to build a plug-and-play product at some point.”
Hasty has taken this feedback on board and created the plug-and-play solution the mining company was looking for with the release of our Model Playground and Inference Engine; find out more about that here.