Build, Deploy, and Secure Predictive AI Models with Databricks
Kumo is a predictive AI solution that accelerates the creation and performance of predictive models from datasets within Databricks Unity Catalog. Kumo combines graph neural networks (GNN) with large language models (LLMs) to learn across all data in the warehouse. Kumo makes highly accurate predictions about segments, lifetime value, behaviors, and more that helps drive business objectives.
Leveraging its internal GNN models, Kumo learns across multiple Databricks tables in customer’s Unity Catalog and builds accurate predictions. The GNN models can be further improved by domain expert models, such as the encoder-based LLMs. With Kumo + Databricks, customers can leverage a diverse collection of LLMs(e.g. DBRX, HuggingFace, and OpenAI to name a few) through Databricks Model Serving to bring general-world knowledge to customer’s relational data via Kumo’s GNN with LLM model.
Kumo’s predictive models can be further refined by domain experts for maximum performance improvements. In hours, Kumo generates batch predictions or embeddings for use downstream.
The out-of-the-box models created in hours by Kumo using graph neural networks and LLMs are up to 30% more accurate than baselines.
Compatibility with a diverse collection of Databricks supported LLM frameworks and libraries offers flexibility in choosing the right tools for specific tasks.
No data is stored on disk in a Kumo owned environment. Data leaving Databricks is transformed, encoded, and deleted after use.
Build your graph once by connecting your Databricks Unity Catalog tables, then use it to generate numerous predictions for various use cases. Kumo’s automated ML pipelines keep models up-to-date, ensuring deep learning from the latest data.
Kumo is built for scale and operates on terabyte-sized tables in Databricks Unity Catalog. There’s no need to sample or reduce data when training, ensuring comprehensive and accurate predictions.
Kumo’s AI learns from the latest data in Databricks Unity Catalog at your chosen interval, ensuring that predictions are always as accurate as possible
Kumo serves real-time model inference or batch predictions, which can be written back to Databricks Unity Catalog or into a key-value store for real-time serving.
Build your graph once by connecting your Databricks Unity Catalog tables, then use it to generate numerous predictions for various use cases. Kumo’s automated ML pipelines keep models up-to-date, ensuring deep learning from the latest data.
Leveraging Databricks Model Serving and its support for LLMs, Kumo combines GNNs with LLMs to learn from both structured and unstructured data in the warehouse, producing highly accurate predictions. Kumo’s prediction results enhance the accuracy of existing models by feeding them trained embeddings, allowing data scientists to add value at each step.
Ensure rapid model deployment and processing while keeping your data secure within your Databricks Unity Catalog. Advanced encoding measures ensure your data and models are protected throughout the ML lifecycle.
