Skip to main content
Choose New project to open the conversational guide. Explain what the model should do, who will use it, and what a good result looks like. The guide asks for missing details and prepares a brief you can edit before saving. Use Use guided form if you prefer to enter the details directly.

​
What to include

  • Task: for example, extract policy numbers and loss dates from claim notices.
  • Model: build a new specialized model, or bring one already trained elsewhere.
  • Data: use a starter or prepare your own supported records.
  • Locations: choose preparation/training and serving destinations separately.
  • Success criteria: define correct outputs, missing-value behavior, and acceptance checks.
  • Budget: record the intended experiment budget. This is a planning value, not a compute reservation or enforced dollar cap.
The guide uses Anthropic for project descriptions. Keep raw documents, personal information, and secrets out of the conversation. Imported datasets are a separate workflow.

​
Review and resume

Review the proposed brief, correct it, and save the project plan. Conversations and project drafts belong to your team. You can return to unfinished setup from Projects. Saving a plan does not train or deploy a model. If someone else edits the same version, the platform rejects the stale save rather than overwriting their changes. Reload the latest version before continuing.

​
Project console

The project list shows the registered model, runtime status, and GPU telemetry when the connected runtime reports it. Missing telemetry is not zero utilization. Admins and developers can edit projects. Members can read them and use a connected model, but cannot edit setup or start and stop demos.