Faced with building their own solution and hiring Dev-Ops staff, Theator found a solution that both saved money and drove productivity.
All surgeries, no matter how long or complex, are deﬁned by a handful of critical moments. Moments where vital, split-second decisions determine whether surgery succeeds, or doesn’t. Where knowing the patient and procedure from all angles is the difference between preventing complications…or creating them. The team at Theator, an innovative up and coming company, is leveraging com-puter vision and machine learning to introduce Surgical Intelligence, a platform that puts deﬁning intraoperative moments in the hands of surgeons so they can continuously perfect their craft. With Surgical Intelligence, surgeons have the technological edge needed to turn precious, data-enriched moments into smarter decisions, sharper skills, and better outcomes.
Hardware Automation and Visibility
At the crossroads of computer vision, surgery, and artiﬁcial intelligence, Theator’s data science team requires extensive computing resources to build, test, and reﬁne their models to incorporate every permutation of each type of surgery they aim to improve. As is the norm in AI development, the data scientists at Theator ﬁrst spend enormous amounts of time creating the initial experiments, then ap-ply the models to massive, complex visual data sets, and then leave them to run while they’re working on modiﬁed versions of that experiment.
After careful analysis, Theator decided that for their needs, it didn’t make sense to purchase a large bank of expensive GPU machines, especially when, aside from peak usage, many would not be in use. “Instead of managing our own hardware for software development”, explains Dotan Asselman, CTO and Cofounder of Theator, “we use Cloud computing so we can spin machines up and down as needed. The downside, of course, is that the costs associated with the time, storage and bandwidth are enormous when we aren’t careful to keep an eye on what is running and when”. To management, allocating dedicated DevOps staff for this seemed wasteful. However they knew they needed a solution to empower their data science team to control and streamline the process and to keep ma-chines running only when they had to be. The problem, of course, was the time and distraction involved in doing so.
As they learned about ClearML, the team believed it had their answer. ClearML is an open-source platform that helps data science and data engineering teams optimize their AI development with a number of tools that yield immediate improvements across four core areas: productivity, collaboration, resource utilization and data management. In less than a week, the Theator team had conﬁgured the ClearML automation and orchestration module, and fully integrated it into their environment. ClearML now automatically spun machines up and down with cost-saving efﬁciency, based on actual demand triggered by the ClearML Agent, a component of ClearML.
ClearML provides an interface to monitor workers and queues that can, on one hand, migrate experiments on demand, and on the other hand, actually execute the experiment…all without requiring the developer to set up the machine, install packages, and rewrite code. As these tasks are handled automatically; Developers no longer need to ﬁrst access a given machine and check its “vital signs” to deter-mine if it has the capacity to manage another experiment.
“A major beneﬁt for us,” explained Asselman, “is that ClearML-agent includes its own machine status reporting engine to monitor GPU, CPU, VRAM, RAM allocation, and net-work IO. As such, we can now trace the impact of speciﬁc code, in a given experiment, on any of these resources. This is a critical data point we often need to discover; when training crashes, we need to know why it happened so we can get back to work. And if my code isn’t maximizing the GPU’s capacity, I’m wasting time and money.”
To achieve this level of transparency through other means, Theator had to either build their own solution, or purchase and manage an external solution. “It feels like ClearML provides a collection of must-have automated process shortcuts that we all knew we needed, but couldn’t, or wouldn’t, create ourselves,” says Asselman.
be around $130K-$170K annually at their current workloads. As their AI workloads grow, so will the savings.
Aside from these savings, Asselmann appreciates the signiﬁcant jump in productivity due to simpliﬁed reproducibility. “For each experiment,” he reports, “adding the ClearML code snippet made the process ‘hands-free’ as developers no longer needed to document every parameter that went into the code and associated packages, then migrate these to a new machine with the data and only then run the experiment. The alternative,” he explains, “means running ‘fairly similar’ experiments without a precise duplication process – which is not only far from best practice, but is also frowned upon in a heavily regulated industry where AI results can mean the difference between success and tragedy. These additional savings are probably on par with the MLOps direct savings, if not more.”
Due to the deployment agnosticity of ClearML, and its ability to run on-prem, on cloud or any combination thereof, Asselmann is conﬁdent that as they grow and potentially invest in their own DGX pods, ClearML will grow with them seamlessly with no change to their workﬂows.
“If I had to sum it up,” concludes Asselman, “I’d have to say that, far from dictat-ing new processes and changing the workﬂows that worked for us, ClearML just feels natural. Quietly and intuitively working in the background, we can focus on empowering surgeons, not wrestling with the steps to get there.”