Build/Buy in MLOPs for R&D Does "off-the-shelf" exist yet?
What kind of tools and infrastructure does a company need in order to build, train, validate and maintain data-based models as part of products?
The straight answer is - “it depends.” The longer one is: “MLOps.”
It is far too early to determine the “best” patterns and workflows for Data-Science, Machine- and Deep-Learning products. Yet, there are numerous examples of successful deployments from businesses both big and small.
Most have done this by building an internal platform to base their R&D on, which enables “productization” of the model-building process. These platforms are scarcely built “from scratch.” Instead, they are dependent on existing hardware, frameworks, and toolchains.
In this webinar, we will try to answer the following questions:
What capabilities should an internal MLOps platform have?
To what extent does one’s platform can be dependent on open-source infrastructure?
Lastly, how “deep” into R&D can one introduce MLOps practices for productivity and reproducibility?
We will be using a real-world example to help ground the ideas, featuring “lessons learned” from infrastructure at theator: A Surgical Intelligence startup company. theator provides revolutionary personalized analytics on surgical operation videos. For that purpose, theator uses continuous training and inference pipelines, with emphasis on straightforward transitions from R&D to “production”. We will also mention topics such as hybrid orchestration through pipeline design, continuous training, and finally, advanced dataset management complying with the privacy requirements.