RosettaHub™ for Data Science

Environments now, surprises never

Start the machine.
Not the conversation.

Data science teams get told no for a reason. A GPU left running is expensive, and a model endpoint nobody is watching is worse. Make those two things impossible and there is no longer much reason to say no.

Described once,
started by whoever needs it

Someone prepares the environment properly one time. After that it is a launch, and it comes up identical every time, which is the part that makes a result worth comparing to last month’s.

Notebooks

Jupyter and RStudio, started from a description so the environment is the same for everyone who opens it.

GPU machines

The instance type you need, with your image and your storage attached, running in your own account.

Spark clusters

The distribution and applications your jobs expect, with the workers on spot capacity where the bulk of the cost is.

Kubernetes

Managed clusters with the storage drivers, ingress and monitoring already wired in, for work that has outgrown a single machine.

The weekend GPU

Everyone has done it. A run finishes on Friday evening and the machine keeps going until Monday, at a rate that would not be approved if anyone were asked.

Idle machines stop themselves

An idle GPU is the most expensive item on a list of twenty eight kinds of waste that are detected across the estate.

Spot without losing the run

Interruptible work runs on spare capacity, and an instance can be hibernated when the cloud reclaims it rather than simply lost.

The budget is checked first

A launch that would take the team over budget is refused by the cloud’s own API, so the limit is real without anyone policing it.

See what gets detected →

Model access

Which models, for whom,
and at whose cost

Most teams discover their AI spend at the end of the month, as one number, with no way to say who it belongs to. Here the model a role may call is a permission, and the tokens it spends are attributed to the person who spent them and drawn against the same budget as their compute, within about five minutes.

That combination is what makes a limit enforceable. A budget running low can narrow which models are reachable, instead of producing a report about a decision that has already been made.

See AI cost governance →

Hand over the environment,
not the instructions for it

Datasets, images, notebooks and whole clusters are shared with a person, a group or an organisation. Access appears when you share and is gone when you stop, across accounts and across clouds, so a machine in one place reaches storage in another without anyone exchanging keys.

See how sharing works →

Common questions

How do data scientists get a GPU machine without waiting for a ticket?

They launch one themselves from a description someone prepared once: the machine type, the image, the storage and the environment. It comes up in your own cloud account inside limits already set, so self-service does not require anyone to relax the guardrails to allow it.

How do I stop a GPU running all weekend?

Idle machines are detected and stopped automatically, and a GPU sitting at nothing is the most expensive example. Work that tolerates interruption can run on spot capacity, where an instance can be hibernated rather than lost, so the saving does not cost you the run.

Can I control which AI models a team is allowed to call?

Yes. Model access is a permission attached to a role, so a team can be given exactly the models their work calls for. Token cost is attributed to the person who incurred it and drawn against the same budget as their compute, within about five minutes.

How do I share a dataset or an environment with a colleague?

Share it directly, with a person, a group or a whole organisation. Access is granted when you share and withdrawn when you stop, including across accounts and across clouds, so a machine in one place can reach storage in another without anyone exchanging keys.