Welcome to prokube

prokube is a data science platform that provides everything you need to develop, train, and deploy machine learning models. Built on Kubernetes and Kubeflow, it offers a collaborative environment where data scientists and ML engineers can work efficiently across the entire machine learning lifecycle.

GenAI & LLM Ready

prokube is purpose-built for Generative AI workloads. Deploy Large Language Models, build agentic applications, and run GPU-accelerated inference at scale.

Capability	What You Get
LLM Serving	Deploy models like Llama, Mistral, and Qwen as autoscaling REST endpoints
GPU Support	NVIDIA GPUs with timeslicing and MIG for efficient multi-tenant usage
Fast Startup	Local model caching for production LLM deployments
Agent Pipelines	Orchestrate agentic workflows with Kubeflow Pipelines
Experiment Tracking	Log prompts, evaluations, and model versions with MLflow
Large Model Storage	S3-compatible storage for model weights and artifacts

What You Can Do

Develop in Interactive Environments

Launch notebook servers with JupyterLab, VS Code, or RStudio directly from your browser. Each notebook comes with customizable compute resources (CPU, RAM, GPU) and persistent storage, so your work is saved across sessions. Pre-configured images include common data science tools, or you can use custom images tailored to your project needs.

Get started with Notebooks

Store and Access Your Data

Access multiple storage options designed for different use cases:

• MinIO - S3-compatible object storage for datasets, model artifacts, and any file-based data
• PostgreSQL & TimescaleDB - Relational databases for structured data and time-series workloads
• Persistent Volumes - Kubernetes-native storage that can be attached to notebooks and pipelines

All storage services are accessible from notebooks, pipelines, and deployed models.

Learn about Data Storage

Build and Automate ML Pipelines

Kubeflow Pipelines let you turn experimental notebook code into robust, production-ready workflows. Instead of manually running cells in order, you define your workflow as connected steps that execute automatically.

Why use pipelines?

Benefit	Description
Reproducibility	Every run uses the same code, data, and environment - no more "it worked on my machine"
Automation	Schedule pipelines to run nightly, weekly, or trigger them from external events
Scalability	Each step runs in its own container with dedicated resources - process terabytes without blocking your notebook
Visibility	Track every run with full lineage: what data went in, what model came out, and every step in between
Collaboration	Share pipelines with your team - they can run them without understanding every detail

When to use pipelines:

• You're running the same workflow repeatedly (daily retraining, batch inference)
• Your workflow takes hours and you don't want to babysit a notebook
• You need to hand off a workflow to someone else or to production
• You want to compare runs across different parameters or data versions
• Multiple steps need different resources (CPU for preprocessing, GPU for training)

Common pipeline patterns:

• Training pipelines: data prep → feature engineering → model training → evaluation → registration
• Batch inference: load model → process data in chunks → write predictions to storage
• Data pipelines: extract from sources → transform → validate → load to warehouse
• CI/CD for ML: retrain on new data → validate performance → deploy if metrics pass

Explore Pipelines

Tune Hyperparameters with Katib

Finding the right hyperparameters can make or break your model. Katib automates this search so you don't have to manually try hundreds of configurations.

What Katib does:

• Hyperparameter tuning: Automatically search for optimal learning rates, batch sizes, regularization, and other training parameters
• Neural architecture search (NAS): Explore network architectures to find the best structure for your problem
• Early stopping: Terminate underperforming trials early to save compute resources

Supported search algorithms: Bayesian optimization, Tree of Parzen Estimators, Random Search, Hyperband, CMA-ES, and more. Katib also integrates with Optuna if you prefer that framework.

Katib works with any ML framework (PyTorch, TensorFlow, XGBoost, etc.) and integrates with your existing training code.

Learn about Katib

Track Experiments with MLflow

Log parameters, metrics, and artifacts from your training runs. Compare experiments, visualize results, and register models for deployment. MLflow integrates with your notebooks and pipelines to provide a complete experiment tracking solution.

Use MLflow

Deploy Models to Production

Serve your models as REST endpoints using KServe. Built-in features include autoscaling, request logging, and A/B testing. Deploy models from MLflow or directly from your storage.

For LLM deployments, prokube supports vLLM and other inference runtimes with GPU acceleration, local model caching for fast startup times, and efficient resource sharing through GPU timeslicing.

Deploy with Model Serving

Getting Started

1. Log in to the Central Dashboard using your credentials
2. Create a notebook to start developing - choose your preferred IDE and resource requirements
3. Explore your data using the pre-configured storage connections
4. Build pipelines to automate your workflows as your project matures

Need Help?

• Check the documentation sections in the sidebar for detailed guides
• Explore the Examples Repository for working code samples
• Contact your platform administrator for access-related questions