4 ms·

We have not found something that solves our problems and is flexible enough, so we built our internal "machine learning platform" (https://iko.ai https://iko.ai

We have not found something that solves our problems and is flexible enough, so we built our internal "machine learning platform" (https://iko.ai https://iko.ai). It's not yet enterprise ready, not by a long shot (no SSO, RBAC, multitenancy, etc) as we're in the process of making it because we need it for yesterday and we're focusing on functionality to actually serve our clients. We have a notebook experience built on top of JupyterLab because in our experience all these years profitably delivering products to large organizations, what slowed us the most wasn't the lack of better notebook stylesheets or animations in UI. We built the following features in our product, and they might be good to look for when you shop around: - No setup collaborative notebooks: collaborative editing on notebooks so you could do group programming and troubleshooting, see people's cursors, etc. Different images that contain the most popular libraries. We'll support fast.ai's fastbook for training. - Long-running notebooks: pretty much fire and forget. It's super useful, I could launch twenty notebooks, see them while they're beeing executed. Log out, close the tab, come back, and see their on-going output. - Automatic tracking for parameters, metrics, and models. We deploy our models with that, too. - AppBooks: enable sharing interactive notebooks with stakeholders without them having to open JupyterLab, or see code.

spicyramen 6y ago
The Long Running notebooks is a feature that we would love to have, we have been using papermill (hacky scripts) and recently I saw a blog post where Google was recommending shipping notebooks to the AI Platform training product as a job which was a good option until something fails is really hard to have our Data scientist look into logs and the context switching between leaving the Jupyterlab UI and going to the Google Cloud console is not good, how you do it?
- Jugurtha 6y ago
  Glad you like it. A lot of work went into focusing on essentials for a team to ship product to paying clients. We're optimizing for that and building what similar teams to ours need to execute real projects. We want at least 10x improvement (time, cognitive load, dreading to do things for fear of breaking stuff which limits experimentation). There are also features you might like, such as automatic experiment tracking (metrics, parameters, and models). You can also deploy the automatically tracked models pretty much in one click and it'll give you a nice "REST" API endpoint to invoke your models. The plumbing for model monitoring is there, we'll add the porcelain very soon to allow users to see it. Currently tweaking data management as well in the "Object Storage" application. We'd love to hear more of your problems. You and your team can sign up for access and start getting work done while you either find a platform that suits you, or we become a platform that suits you. It will be sooner than you think, as we've done a lot of - non user facing - plumbing recently.

spicyramen 6y ago

The Long Running notebooks is a feature that we would love to have, we have been using papermill (hacky scripts) and recently I saw a blog post where Google was recommending shipping notebooks to the AI Platform training product as a job which was a good option until something fails is really hard to have our Data scientist look into logs and the context switching between leaving the Jupyterlab UI and going to the Google Cloud console is not good, how you do it?

Jugurtha 6y ago
Glad you like it. A lot of work went into focusing on essentials for a team to ship product to paying clients. We're optimizing for that and building what similar teams to ours need to execute real projects. We want at least 10x improvement (time, cognitive load, dreading to do things for fear of breaking stuff which limits experimentation). There are also features you might like, such as automatic experiment tracking (metrics, parameters, and models). You can also deploy the automatically tracked models pretty much in one click and it'll give you a nice "REST" API endpoint to invoke your models. The plumbing for model monitoring is there, we'll add the porcelain very soon to allow users to see it. Currently tweaking data management as well in the "Object Storage" application. We'd love to hear more of your problems. You and your team can sign up for access and start getting work done while you either find a platform that suits you, or we become a platform that suits you. It will be sooner than you think, as we've done a lot of - non user facing - plumbing recently.