Skip to content

Choosing Your Workflow

PAIS supports two distinct usage patterns. Most researchers use both at different stages of a project.

Use case 1: Inference consumption

Call hosted AI models from your own code.

Your scripts, notebooks, or agentic frameworks call the PAIS API over HTTPS. You don't need kubectl, a namespace, or any platform-level access beyond an API key.

Your code  →  PAIS API (OpenAI-compatible)  →  Llama / embedding model

Best for:

  • Literature review and summarisation pipelines
  • Embedding documents for semantic search
  • Agentic research assistants (LangChain, LlamaIndex, Claude Code)
  • Generating code or analysis in a notebook
  • Any task that treats the LLM as a black-box API call

What you need: An API key. That's it.

Get started with the API →

Use case 2: Active compute and development

Work directly on the platform with GPU access.

You get a namespace on the cluster with GPU quota, network-attached storage, and a Kubeflow environment. You develop and run AI training jobs, pipelines, and experiments directly on the platform.

Jupyter notebook  →  GPU (L40S, time-sliced)  →  Training job
                                                 ↘  Kubeflow Pipeline
                                                 ↘  MLflow experiment

Best for:

  • Training or fine-tuning models
  • Hyperparameter sweeps (Katib)
  • Automated ML pipelines (Kubeflow Pipelines)
  • Working with large datasets on shared VAST storage
  • Developing inference code that will eventually be hosted on PAIS

What you need: Research group membership + Kubeflow access provisioned by the PAIS team.

Get started with notebooks →

Combining both

A typical research software workflow uses both:

flowchart TD
    A[Explore with Jupyter\n+ GPU notebook] --> B[Prototype pipeline\nwith Kubeflow]
    B --> C[Fine-tune model\non VAST data]
    C --> D[Register model\nin MLflow]
    D --> E[Serve via PAIS\nInference API]
    E --> F[Call from agentic\nresearch tools]

Start with Use case 1 (API calls) while your platform access is being provisioned — most agentic development work only needs the API.