SciFlow
Launch reproducible GPU environments, scale training, compare experiment results, and work with an agent that understands your infrastructure.
Built for labs running private or on-premises GPU clusters.
Keep environments, training runs, experiment results, and agent context connected from exploration to evidence.
Reproducible work
Launch a maintained GPU environment with project storage and access already wired. When the environment works, preserve it as an image so collaborators and future runs can begin from the same state.
The result keeps its context.
Code, dependencies, data mounts, and compute history remain connected to the work instead of being scattered across setup notes and messages.
Scalable experiments
Describe the workload, choose protected or reclaimable capacity, and follow every rank from admission through completion. SciFlow turns shared GPUs into a predictable research instrument.
Waiting is explainable.
Researchers can see why a run is waiting, what resources it holds, and where it failed without chasing the platform team or reading Kubernetes events.
Experiment tracking
Track loss, accuracy, and other experiment metrics alongside run configuration. Compare runs in SciFlow and follow each result back to the instance, job, or distributed training run that produced it.
Start logging without the setup.
Use supported wandb logging calls or the Hugging Face integration in SciFlow workloads. No tracking package to install and no API key to manage.
Research agent
The built-in agent sees your workspace and the live SciFlow API. It can inspect runs, reason over project context, create visual artifacts, and execute infrastructure actions within your permissions.
Assistance stays inside policy.
API access follows your user and organization permissions. Mutations follow the selected approval policy: pause for review or run automatically. Platform authorization applies in either mode.
SciFlow connects experiment signals with the policies that govern shared compute, so researchers can resolve problems independently and plan work around real capacity.
01Observe the run
Follow utilization, rank health, logs, events, scheduling state, and cost without leaving the experiment.
02Explain the wait
See whether work is running, provisioning, or queued—and which resource or policy is shaping the decision.
03Plan the next launch
Review available quota and launch impact before committing scarce GPU capacity.
Workload observability
Workload telemetry, rank state, filtered logs, scheduling history, Kubernetes events, and usage cost stay attached to the same research object—from a quick workspace check to a multi-node training run.
What this changes for researchers
Researchers run their own work, lab administrators shape access and budgets, and platform teams retain control of policy and GPU supply.
Transparent quotas
SciFlow turns lab policy into information researchers can act on: guaranteed and shared GPU capacity by type, member-level limits, fractional GPU allocation, and CPU, memory, and storage boundaries.
What this changes for labs
Where we are going
To help run science, an agent needs governed access to compute, files, telemetry, and experimental state. SciFlow is building that operating layer.
Today
The agent can inspect the lab’s live environment, reason over workspace context, and carry out platform work under your permissions and selected approval policy.
Direction
The same foundation can support agents that coordinate environments, execution, monitoring, and analysis while scientists set intent and review evidence.
Early access
We’re working with academic teams running private GPU clusters. Tell us where infrastructure slows your science down.
