An advanced engineering team looking for a shortcut to running a model on specialized processors
[redacted metallurgy company] manufactures machines for the quality control of advanced materials, cutting materials, metal grinding, and polishing to prepare them for microscopy. They help companies by providing industrial equipment to perform this advanced analysis. To that end, they have to ensure that the correct process is followed to inspect these materials. They are used by manufacturers worldwide.
Their software development team was tasked with adding vision AI to the machines they produce to ensure the correct execution of the process they enable. The team needs to assess the amount and type of material inspected to inform the machine’s operations.
The machine they built had a specific low power and low heat GPU installed to meet the other engineering requirements. This made deploying the model to this hardware non-trivial, as the low-level software coordinating operations on those processors did not support modern model architectures. They needed to run modern models to reach the performance levels they demanded from the application.
While the team was very technical, machine learning was new to them. They quickly clued themselves up on the state of the art in vision AI and made two key hires.
They are not only the go-to guys for annotations – we have never seen a better tool – but we also felt that they would be a perfect collaboration partner – another task that fits. They have a good, iterative process that keeps getting better.
Software Coordinator and Architect
Forming a partnership with a young startup
Their embedded developer found Hasty and tried out the free trial. So impressed with the automatic annotation, he decided to use the standard algorithm provided by Hasty to train his first neural network. They used this to prove that modern models could do what they needed.
Now they needed to retrain the model for their specific problem. They approached the Hasty team stating that they required algorithms for detecting specific features for microscopy pictures and samples.
Firstly, they needed to label more images. For this, they required high levels of automation in the annotation process as they could not outsource this work. This was partially due to the company’s security policies, but mainly because only their internal experts understood the relevant information in the images.
Second, they wanted to implement their data flywheel on this relatively unique GPU and needed support to get the models to run on this a-typical hardware. They had internal pressure to prove that this technology would work for their use case on time not to delay the overall launch of the machine. They needed to be fast.
Models running on machines in the market
The benefit for [redacted metallurgy company] was to go from the raw image to having the model they needed for their machine from a single partner. “The application was easy to work with, with a very professional look. It worked out of the box and, with the updates, became better and better. Running ML models becomes even more manageable. By now, there is a robust software base that leaves an excellent impression.”
“Without Hasty, the annotation work would not be as fast and accurate. Originally, we thought about hiring people to annotate everything, but it was so fast that our experts could do it themselves. Secondly, without Hasty, we would have needed some type of help from a consulting firm to develop algorithms to meet our demands.”
“In the long run, the team will maintain and take ownership of the design details of the model. We can do all of this on the Hasty platform. We would not like to depend on some service provider to put something in our product like a model. We don’t mind paying for the model building phase – but once it’s developed, we are not interested in dependencies on services.”