Client Overview: Ultraleap
Ultraleap is a computer vision company creating hand tracking machine learning models for cutting edge VR and XR devices. They train over 90 variations of hand tracking models for different devices, SoCs and operating systems. Managing the data, training, deployment and subsequent lineage for all of these models, particularly with the rate at which their research team works, is incredibly challenging. ClearML’s end-to-end machine learning platform has assisted them greatly as a component of their MLOps platform.
We recently caught up with Sam Jenkins, Ultraleap’s MLOps Technical Lead, to discuss how they use ClearML’s end-to-end machine learning platform for managing the lineage of training and deploying a large number of models.
The Challenge: Manage Complex Workflows
One interesting aspect of Ultraleap’s machine learning workflow is that they simulate significant amounts of data to train their hand tracking models, however this adds complexity to their already intricate experiment lineage. When running new experiments their team rapidly generates new synthetic datasets, runs data augmentations, trains models, and packages the models for multiple SoC and OS types. With so many moving parts, Ultraleap needed a tool that could easily manage the complexity of maintaining these variables in their workflow, and they chose ClearML’s solution for its experiment management capabilities, data version control, and workload orchestration.
The Solution: ClearML for the Win
ClearML’s orchestration system integrates into Ultraleap’s Kubernetes cluster and allows them to run data simulation across multiple nodes in their cluster. ClearML automatically tracks the configuration for those simulation runs, along with their generated data, which saves time manually logging all the configurations and generated data.
ClearML’s open source API allowed Ultraleap to easily integrate to their GitLab pipelines for packaging and deploying models for testing and deployment on varied hardware targets.
ClearML’s data management system allows Ultraleap to track, manage, and share new datasets for future experiments’ usage. They maintain a number of “baseline” datasets and iterate on top of them to change various parameters, e.g. lighting conditions and backgrounds, effectively building data lineage from “baseline” datasets.
ClearML gives Ultraleap flexibility around how they manage their compute resources, making it easy to manage both large-scale cloud infrastructure and small-scale on-premises hardware.
The Results: Faster Research and Time to Market
The ability to quickly go from deployed model back to dataset, and all data processing in between, has significantly sped up the research team’s work, and has enabled faster throughput from ideation to product. Integrating Ultraleap’s downstream software packaging processes using ClearML’s API has also allowed better cross-team delivery of Ultraleap’s machine learning models. ClearML’s UI and experiment tracking also allows their team to easily visualize video data and build custom metrics handlers to support ease of comparison between datasets and models.
Editor’s Note:
Ultraleap uses ClearML’s AI Development Center, an open-source solution that streamlines AI/ML workflows from experimentation to production. With minimal setup, developers can efficiently build, train, and deploy models at scale while leveraging automation, orchestration, and dynamic compute resource management. Seamlessly integrating with ClearML’s Infrastructure Control Plane, it ensures secure execution across diverse environments, including on-prem, cloud, and hybrid setups. The platform enhances collaboration through structured experiment tracking, federated data management, and simplified model deployment via CLI, UI, or API. Designed for scalability and enterprise security, ClearML’s AI Development Center helps teams accelerate AI innovation while reducing operational complexity. To see ClearML in action, please request a demo to speak with our sales team.