web scale online learning,
simplified

Get an always-on, scalable inference endpoint for your model in minutes, update the model weights in milliseconds

Register for an account

The Simple Path from Model to Production API

Your data pipeline can retrain a model in seconds, but pushing it live is slow. diagonal.sh provides the missing piece: an API that makes model deployment and updates an atomic, millisecond operation. Here’s how it fits into your existing stack.

1

Train Your Model as Usual

Use scikit-learn to train or retrain your linear modelin your existing environment. You retain full control over your infra.


    # Train in your existing notebook or pipeline
    from sklearn.linear_model import LogisticRegression

    model_v1 = LogisticRegression()
    model_v1.fit(X_train, y_train)
2

Deploy with One Command

Use our Python package to deploy a scalable, secure, and production-ready API endpoint for your model.


    # Get a live API endpoint instantly
    from diagonal import deploy

    endpoint = deploy(model_v1) # API key is set via env variable
    # deployment_metadata["model_id"] == 'i312jjse...kajkdu'
3

Update Atomically

When your data pipeline computes new weights, just call `deploy` again with the same `model_id`. We handle the hot-swap instantly with zero downtime.


    # In your Kafka consumer or Spark job...
    model_v2 = retrain_model_on_new_data()

    # Update the live model in milliseconds
    deploy(model_v2, model_id=deployment_metadata["model_id"])
4

Use from Anywhere

Your high-throughput REST API is ready to be integrated into any application. It's that simple.


$ curl -X POST https://infer.diagonal.sh/$ROUTE \
        -H "X-Api-Key: $API_KEY" \
        -d '{"data": [5.0, 3.1, ...]}'

Unopinionated and Universal Integration

diagonal.sh doesn't care about your stack. It's a simple, universal API endpoint designed to plug into any environment that can run Python. Your training logic can live anywhere—from a data scientist's laptop to a massive streaming cluster. The integration point is always the same simple, powerful API call.

💻 Local Development (Jupyter, VS Code)
🗓️ Batch Job (Airflow, Databricks)
⚡️ Streaming Pipeline (Spark, Kafka)
→

The Universal Integration Point

# The only line you need to add
deploy(model, model_id)
→
🚀 Stateless Inference API
model_v3.14_final_final
📲 Your Application (Makes a REST API Call)

Level Up Your ML Products

Detect Fraud

(FinTech/Bank/Payment Provider) Classify transactions as legit or potentially fraudulent on the fly, and react rapidly to changing patterns of misuse.

Identify Bots

(Gaming Company/Customer Facing Platform) Classify clients as bots based on usage profiles, and maintain the integrity of your game or platform. Roll out each new model as soon as it's available.

Manage Financial Risk

(Financial Market Participant) Predict near-term market volatility using real-time price and volume data. Update your models as quickly as market conditions change.

Counter Cybersecurity Threats

(Cyber Security Firm) Update the models that identify malicious network traffic as quickly as the data comes in. Get the new model live the moment it has been trained.

Query Routing for LLMs

Route queries to the AI System, Agentic Workflow, Human-in-the-loop Workflow or System (like a ticketing queue) that is most likely to perform well (on a risk- and cost-adjusted basis).

Learn Adaptive System Prompts for your LLM App

Construct the system prompt on the fly, so that it optimally serves each user query.

Dynamically set LLM Inference Hyperparamers

Set the inference hyperparameters (such as thinking budget) on the fly, and adapted to the user query.

Counter Concept Drift in Computer Vision Models

Process each incoming image to an embedding via a static model deployment, then use a dynamic endpoint to do the final classification or regression. Never let your models go stale again.

Adapt Your NLP Models to New Data

Embed the text via a model provider, then use a dynamic endpoint for classification or regression. Now you can easily adapt to changing patterns in the data, and even add or remove categories that you need to identify.

Online Learning on Custom Embeddings

You can't just embed text or images, you can embed anything you want. Implement prediction, decision-making or monitoring models that adapt to new data on your custom embeddings.

A Tool for Streamlined Model Deployment

Atomic Millisecond Updates

Decouple your learning frequency from slow deployment cycles.

Spec: Update model weights in less than 50ms

Decoupled & Unopinionated

diagonal.sh is compatible with any data architecture, processing technology or MLOps platform that can run Python.

Spec: Export from any Python process, in the cloud, on prem or on your laptop

High-Performance Serving

Built for web-scale traffic with predictable low latency.

Spec: Sustains 10,000+ requests/sec per endpoint

Secure

Focus on your logic, knowing the infrastructure is secure.

Spec: Secure endpoints authenticated with api keys you issue and revoke

Simple, Transparent Pricing

Total Cost = Subscription + #Models + #Requests + #ModelUpdates

Subscription Tiers

Shannon

€30/month

Run your first online learning application

  • 3 Routes
  • Unlimited models per route
  • Community support
Get started

Noether

€100/month

Expand to multiple usecases

  • 10 Routes
  • Unlimited models per route
  • Community support
Get started

Kolmogorov

€300/month

Wide & deep integration into your service

  • 30 Routes
  • Unlimited models per route
  • Priority Support
Get started

Metered Billing

€10
per 1 million requests
€0.01
per model per day
€0.001
per model update

All prices are before value added taxes, where they apply.

Calculate Your Monthly Cost

Calculated costs are before value added taxes, where they apply.

Meet the Team

Team Member 1

Leon Luithlen

founder and ceo

After 7 years in the data science/machine learning industry, Leon founded diagonal.sh to make machine learning deployment as easy as possible.



Our Purpose

Machine learning can enable us to make better decisions, and help us cater to the needs and requirements of each of us. We aim to make model deployment as easy and seamless as possible, to empower more people to build ML products and enable the spread of ambient, pervasive intelligence, so that humanity can flourish.

Welcome to the future of machine learning.

Register for an account