[redacted startup] has built a hand scanner that looks for contamination on people’s hands. The first use case was to ensure that doctors and nurses had cleaned their hands effectively before surgery.
They were a small technical team that needed to quickly deploy the vision AI models to their low-powered mobile devices. The goal was to improve the accuracy of the model. They are a team of senior software engineers who were relatively new to machine learning.
The team initially started with GIMP, a standard image editing tool, but had difficulties “farming out” the labeling work. They then turned to Amazon’s Mechanical Turks – with lots of frustration. Then, they found Hasty and realized they could get from image to model with their team in a single platform thanks to the model-assisted annotations and licensing the models directly from Hasty.
Because of Hasty, [redacted startup] has been able to accelerate the development of key features. Open communication and clear dialog with the team have allowed our engineers to focus. The rapid iteration and strong feedback loop mirror our culture of a fast-moving technology company.
Senior Computer Engineer in charge of ML
As Hasty doesn’t re-use the images for downstream customers, they were confident that they could serve their current application using their data set and still keep proprietary ownership. Therefore, they could build their application while securing their competitive advantage.
“Our images are really, really low quality – two-channel images that are noisy. The challenge was to trace around someone’s hand perfectly; this problem could not be solved with traditional rule-based algorithms. With machine learning, we were able to solve this problem.”
When they tried to get the segmentation work done manually, the price point was too expensive. Furthermore, the privacy side made it near impossible to outsource this work. One has to make the images publicly available, making them vulnerable to spiders searching images, etc. This solution had insufficient data privacy or security.
They even thought about building their own tool like Hasty to farm out segmentation but realized this would be way too much work. Together with Hasty and a trusted annotation service team, they created the ground truth dataset needed to get the application shipped.
“The value comes in the model training – having the model training in the same environment that we are labeling the data in means that the connection between input and the desired output is really close. We were also impressed how easy it was to bring new team members into the platform.”
[redacted startup] has now deployed a market-leading solution on time and within budget for their primary use case. They have generalized this application to work at multiple locations reliably and look at adjacent use cases on the back of this success.