MLOps at Ultraleap
— 1 min read
(https://www.youtube.com/watch?v=_Fqzd1dldYs)
I was hired at Ultraleap to replace a legacy MLOps system that was no longer maintainable, known as "Generics". Part of the process of replacing the system was to evaluate a lot of technologies against the existing system, and see what the best fit would be for the team. Given I was a one man band in the MLOps team at the time, I had the age old conundrum of balancing off the shelf tooling with custom software. Some of the questions I had to ask:
- How do the features of new off the shelf MLOps tooling compare against the legacy system? The system was very well loved, and was built on over years at LeapMotion (which merged with Ultrahaptics to form Ultraleap).
- How will the team adapt to a new way of working? and how can we ease that transition?
- What are the "non-negotiable" features that we require?
- How do we develop and maintain our backend machine learning infrastructure on AWS?
I was lucky enough to be invited to talk about this in the Bristol MLOps meetup group, which Ultraleap hosted back in 2022. Here I talk through the above questions, and also show some of the very cool tech that we work on at Ultraleap! The video is here !
NB - spoilers, we did choose ClearML, with our own backend infrastructure on kubernetes, which I'll be writing about in future blog posts.
