# Feature store and MLOps

> DataNature AI gives data science teams a feature store, vector database, training, experiment tracking and MLOps to deploy, version, monitor and serve models in real time or batch. Models feed AI360 Studio decisions and Voir agent skills, with lineage back to training data.

Source: https://datanature.ai/solutions/feature-store-mlops
Publisher: Dlytica Inc. (DataNature is a Dlytica company)
Updated: 2026-11-05

From notebook to governed production model

## Components

- **Feature store**: Reusable governed features
- **Experiment tracking**: Runs and metrics
- **Model serving**: Real-time and batch endpoints
- **Atlas**: Model registry, lineage and logs

## Steps

1. **Engineer features**: Build features from Gold tables with Spark.
2. **Train and track**: Compare runs and pick a model.
3. **Deploy**: Serve in real time or batch.
4. **Monitor**: Watch drift and performance; retrain on schedule.

## Documentation

- [Machine learning](https://docs.datanature.ai/docs/developers/ai-and-analytics/machine-learning)

## Can we use our own frameworks?

Yes. Standard Python and Spark ML tools run on DataNature compute.

## Where do models run?

Inside your deployment, including open-weight LLMs.

## How do models reach the business?

As AI360 Studio scores and as Voir agent skills.

## Common questions

### What is a feature store?

A shared, governed library of model inputs with the same definitions for training and real-time scoring.

### Which features work for fraud models?

Velocity, deviation, newness, network and context. See [the fraud features playbook](/blog/fraud-detection-features-playbook).

### How do we meet model risk rules?

Inventory, lineage, reproducible training and monitoring. See [OSFI E-23 explained](/blog/osfi-e-23-guideline-canadian-financial-institutions).

## Implementation details

1. **Build features in dbt Gold** from certified data, and slow features on Spark for large histories.
2. **Add real-time features** from StreamHouse for models that score events (fraud, alerts).
3. **Train in notebooks** on the shared Spark server; log runs, parameters, metrics and artifacts to DataNature AI experiment tracking.
4. **Register models** in the model registry with versions, aliases and lineage to the training run.
5. **Serve** for real-time or batch inference, and write every prediction back to Iceberg with the model version.
6. **Monitor** drift and performance against outcomes, and retrain from the same Gold features.
