What did you think was going to be necessary but turned out not to be?
What are you still looking for?
What's almost right but not quite?
We do machine learning and it just sucked to see people stress about finding transportation (bus), traffic jams, and suffer unreliable trains.
This coincided with our effort to capture the experience we acquired doing several machine learning projects to build our MLOps platform[0]. The first goal of this was to make it possible for us to train models from home, without having to come to the office to use the proverbial "powerful work station" in every small ML shop.
Then we added fresh environments from images to reduce the number of times our colleagues who had a more academic profile needed us to set up their compute environments. Then we added real-time collaboration in notebooks to make pair programming or troubleshooting easier. Then we added scheduled long-running notebooks as opposed to one person telling the other they're launching a trainin job, etc. Then self-service deployment. So we built our platform to enable us to work remotely on ML projects.
We do calls every week or two weeks to talk through some issues. We do that on Jitsi and have a collaborative notebook on our platform where there's the agenda, and on which we can add code snippets demonstrating bugs, or what would a solution look like. We record the videos and make them accessible to everyone so they could refer to them later if necessary.
We struggle with lousy internet connection. For some of us, it's really, really, bad. That, however, was an issue even when working from the office.
Overall, it's really working for us.
- [0]: https://iko.ai
It is very open with minimal features which means less confusion and good adoption.