Template-backed instances
Launch long-lived SSH, Jupyter, or custom entrypoint environments from reusable templates with versioned defaults.
Early access
Launch research GPU workspaces without fighting the cluster
SciFlow gives labs and AI teams a self-service control plane for template-backed GPU instances, org quotas, image commits, and usage visibility on top of Kubernetes.
- Built on
- Kubernetes
- Control plane
- Rust services
- Identity
- OIDC / Authentik
Launch instance
PyTorch + Jupyter
Starts now
H100
A100
L40S
Templatepytorch:2.4 + jupyter
Org chargedVision Lab
GPU shape1 × H100 (1/1)
Auto-stop12h
The problem
Research compute should not depend on Slack messages and manual kubectl
Most teams build the same control plane twice — once in scripts, once in tickets. SciFlow replaces both.
GPU access is hard to share fairly
Without a real admission layer, the loudest user wins and queues become Slack threads.
Users want environments, not raw pods
Researchers need long-lived SSH or Jupyter workspaces, not short-lived containers tied to a pod spec.
Image save and reuse is manual
Saving a working environment for the next experiment usually means scripts, registries, and copy-pasted commands.
Admins lack quota and usage visibility
Org-level quotas, member quotas, and GPU usage rollups end up scattered across spreadsheets and dashboards.
Product
A control plane for interactive GPU research
Six product surfaces that together replace the patchwork of scripts, dashboards, and manual approvals most clusters rely on today.
Org-aware quota
Users pick which org they are charging while admins manage member quotas and integer GPU budgets per type.
Queue-first admission
Reject invalid requests, queue valid ones when capacity is temporarily unavailable. No empty pods reserving GPUs.
Commit running work into images
Save a configured workload as a reusable image to seed future templates and reproducible experiments.
Account-scoped keys
Keep SSH keys and API keys attached to accounts and injected at launch time — not buried inside templates.
Usage and reporting
Usage records, billing summaries, GPU accounting, and scheduled rollups for cluster admins and finance.
Workflow
From idea to running GPU workspace in minutes
A consistent path from template selection to a saved, reusable image — the same loop researchers already follow informally.
- 01
Pick a template
Choose a versioned template — image, launch mode, startup script, ports, env vars.
- 02
Choose GPU shape and org
Pick the GPU type and fraction, and select which org membership you are charging.
- 03
Launch or queue automatically
Admission runs quota math. Start now if capacity exists, otherwise queue fairly.
- 04
Save the environment as an image
Commit your running workload back into a reusable image for the next experiment.
Admin · Quotas
Cluster overview
live
GPUs total
24
In use
11
Queued
3
| Org | H100 | A100 | Queue |
|---|---|---|---|
| Vision Lab | 4 / 6 | 2 / 4 | 1 |
| NLP Group | 1 / 2 | 3 / 3 | 0 |
| Robotics | 0 / 1 | 1 / 2 | 2 |
For admins
Built for platform teams, not just individual users
SciFlow gives platform admins the controls and visibility they need without forcing the underlying cluster into a rigid layout.
- Create orgs and appoint org admins
- Set org GPU quotas per gpu_type
- Set member-level quotas in 1/8 GPU steps
- See queue state and runtime status
- Keep Kubernetes as a generic execution pool
- Avoid hard-binding physical nodes to groups
Architecture
Designed as a Rust control plane for Kubernetes
Five focused services with clear ownership boundaries. Authentik handles login at the edge; SciFlow handles authorization, quota, and runtime.
P
Policy
Local user projection, quota ledgers, and admission decisions live here.
O
Orchestrator
Templates, instance lifecycle, queue state, and runtime status.
I
Images
Image metadata, commit operations, registry ownership.
O
Operations
Durable workers, retries, reconciliation, node-local execution.
R
Reporting
Usage records, billing summaries, GPU accounting rollups.
Outside SciFlow
Identity, ingress, Postgres
Authentik, oauth2-proxy, and cluster infra are managed in a separate FluxCD repository — SciFlow stays application-only.
Who it's for
For teams that run serious research infrastructure
University AI labs
Shared GPUs across students and projects with sane quotas.
Internal ML platform teams
A self-service surface that replaces ticket queues.
GPU cluster administrators
Org and member quota math without empty placeholder pods.
Startup research teams
Reproducible templates and saved images per experiment.
Shared compute environments
Org-aware fairness with explicit queueing and leases.
Bring self-service GPU compute to your research cluster
SciFlow helps teams launch, govern, and reuse interactive GPU environments without turning Kubernetes into the user interface.