"Kumo is a predictive modeling platform that empowers data scientists to quickly build highly accurate models. Kumo supports useful, proven AI scenarios like recommendations, personalized communications, next best action, and more. Using the entirety of the data warehouse, composed of multiple tables of structured and unstructured data, Kumo uniquely combines pre-trained LLMs for processing textual and image data with graph transformers. This produces highly accurate predictions and eliminates feature engineering and training set generation. Kumo accelerates predictive model creation from months to days, and improves model performance by double digits.
Kumo uses pre-trained LLMs to process textual and image data and then creates a graph where each row in the warehouse is represented as a node and primary-foreign keys as connections. Then, it applies graph transformers and attention architectures to learn how to combine all the information to produce a prediction. This process, embedding tuning, is where foundation model embeddings are transformed using tabular data in data warehouses to achieve the best performance for predictive tasks.
"