Core
Registry, deployment and monitoring inside your own infrastructure.
Register every model as a versioned artifact, deploy it to compute you host, and monitor it where it actually runs. Open source, framework neutral, fully reproducible, and no inbound access to your network required.
Capabilities
From experiment to production
Registry
A central catalogue where models, experiments and datasets are stored as self-contained .luml artifacts, versioned and grouped into collections.
- Version models and promote them from staging to production
- Store artifacts in your own object storage
- Deploy .luml artifacts without repackaging
| Name | Type | Tags | Created | Size |
|---|---|---|---|---|
churn_clf Id: 019e...ce1f | Model | tabular+2 | ||
demand_forecast Id: 019e...42b7 | Model | time-series+1 | ||
customers_26 Id: 019e...8d50 | Dataset | quarter_1 | ||
customers_27 Id: 019e...b6e3 | Dataset | quarter_4 |
Experiment snapshot
The training run that produced a model, its hyperparameters, metrics and logs, preserved inside the artifact itself.
- Inspect the exact configuration behind any registered model
- Compare several model versions side by side
- Reproduce a past run at any time
Model parameters
Dynamic metrics
rmse
by iteration
Lineage
A graph of an artifact's origins and dependents: the dataset version and experiment behind it, and the deployments running it today.
- Trace a prediction back to the data it was trained on
- Trace forward from a dataset to every model built on it
- Follow the chain across experiments, artifacts and deployments
Deployment
A registered model running as a service on a Satellite - a self-hosted compute node that polls outward for work, so no inbound network access is required.
- Deploy any registered artifact in one click
- Serve real-time inference endpoints or run batch scoring jobs
- Keep artifacts and inference data inside your own network
Endpoint
$ curl -X POST "$LUML_URL/compute" \ -H "Authorization: Bearer $LUML_TOKEN" \ -d '{"inputs": {"price_index": 4.20}}'
Monitoring
A continuous check on a deployed model, running on the Satellite that serves it.
- Detect data drift, feature drift and prediction drift
- Break results down by segment, region or cohort
- Validate schema and data quality on live traffic
Feature drift · PSI
Architecture
Every connection starts on your side
LUML is a control plane: it coordinates work and stores metadata, while models run on Satellites hosted inside your own network. A Satellite polls LUML for work, pulls artifacts from the bucket you connect, and computes its metrics on the node it serves from. There is no inbound path.
Why teams choose LUML
Security and control, by design
Open source, Apache 2.0
The full platform is available under a permissive licence, with no commercial dependency.
No vendor lock-in
Artifacts live in your own object storage in an open, documented .luml format.
Secure by architecture
The Satellite polls outward for work, so no inbound network access is required.
Production-ready by design
.luml artifacts carry everything needed to run, so deployment to a Satellite takes one click.
Granular access control
Control who can view, edit, or deploy across projects. Permissions at the team and workspace level.
Artifact lineage and provenance
Trace any model back to the experiment, code, and data that produced it. Know exactly what’s running in production and where it came from.
Get started
Up and running in minutes
Create a free account
Sign up and get access to the full platform. No credit card, no trial limits.
app.luml.ai/sign-upConnect your storage
Attach any S3-compatible bucket or Azure Blob Storage. Your data stays in your cloud, LUML never stores it.
Settings → Buckets → New bucketCreate your first model
Register a model through the UI, or track your first run with the Python SDK.
pip install luml-sdkFAQ
Frequently
asked
questions
A self-contained model file: weights, metadata, dependencies, and preprocessing in one place. tracker.log_model() creates one from scikit-learn, XGBoost, LightGBM, CatBoost, or LangGraph objects.