How machine-learning startup Jemsoft turned a tragic situation into a viable business

One Monday afternoon in April 2013, 19-year-old Jordan Green was working in a liquor store in Adelaide, Australia, when two men in balaclavas holding a shotgun entered the store, jumped the counter, held the gun to his head, and demanded his co-worker open the store's safe.

It isn't the typical foundation for a company, but this is how Jemsoft was born.

As a pragmatist, Green told ZDNet that he approached the situation by questioning how they entered the store with automatic doors and security cameras. Fortunately, Green was also a programmer involved in robotics.

"The question in my head was why is it that someone who so clearly was not here to grab a slab could come into the local bottle-o and threaten my life and the life of my co-worker -- who to my knowledge has not returned to work. You could say that I took a pretty radical career change because I then left uni, left that job, and tried to build a company, which is not something a sane person would do," he said.

"The reason for that was I felt it shouldn't be that hard to build a system that can identify at the very least that somebody is wearing a balaclava and carrying a gun, and perform the inference that they're probably not there to grab a slab -- especially when I knew I had the technology available."

Studying computer science and already working on a number of projects with Jemsoft co-founder Emily Rich, Green went to Rich with the idea of developing a platform that could prevent situations like the one he had found himself in.

The pair spent a couple of months looking at the space and the viability of the product, and decided to put together a prototype.

They then found themselves a couple of advisors, put a board together, filed for a patent, and started building the system -- only to find the product's market balked at the cost.

With deep learning not as mainstream three years ago, Green explained that hardware manufacturers weren't building edge devices for what Jemsoft needed, so he and Rich were simply "hacking things together" to make something work.

"The hardware we needed for this was hugely expensive, which meant we had an expensive solution. It fell between AU$5,000 to AU$10,000 a year to use -- which we felt was not too much to protect people from having guns held to their heads -- but corporates unfortunately put a lesser price than that on it," Green said.

It became clear that a company's insurance covered the money side of a robbery, and that was the bottom line.

"It's the typical story of solving your own problem. It's an incredible thing in a startup to solve your own problem because the fire of passion will never dwindle, but it also means there are blinkers on when it comes to the business viability," Green said.

"When we came to that realisation, it was very, very hard."

In building the product, Green and Rich had to annotate hundreds of thousands of images and videos of armed hold-ups, which sometimes involved playing dress-ups and reenacting robberies.

Green built a system that allowed a non-technical user to use it, and then decided to upload it to oDesk -- now Upwork -- to outsource it. However, Green said something clicked and made him stop the upload.

"So we stopped uploading the software, because that may have been worth more than anything else we had," he explained. "What we realised at that point is what we had built was not so much a security system, but an entire platform to enable the security system. And that was the gold."

When the duo worked that out, they put an API on it, and MonocularAPI was the result. The computer vision API platform was sent to the cloud and Jemsoft started making sales -- just shy of AU$1 million during the 2015-16 financial year.

Jemsoft started on Microsoft's Azure cloud, but lacked the GPU power that made Green and Rich's life easier.

The pair was contacted by a systems architect who worked with Amazon Web Services (AWS), who put them in touch with some of AWS' hardware designers in Seattle.

With a combined age of less than many of their competitors, Rich and Green found themselves involved in working on what the future of the Amazon GPU cloud would look like.

"For us this was fantastic, because we knew for us it meant that our business trajectory was going to be supported, and on their side I'd presume they were hoping whatever we're doing now, other people would also be doing," Green explained.

"So we were very fortunate to do that, and as a result we were able to design solutions around products that weren't released by Amazon yet.

"This allowed us to be the first general availability -- we beat IBM Watson to general availability in the computer vision space."

In order to do what its customers wanted, there were only two options for Jemsoft: The first was to send customers' data from the Jemsoft software to an AWS datacentre; and the other was for customers to send their information digitally or on a hard drive to Adelaide, for the company to then load it onto physical computers and process it on their own GPUs.

The pair scoffed at the viability of the latter.

At 23 and 26 years old, respectively, Green and Rich have built a powerful and extensive computer vision framework for the cloud.