Berlin Tech Meetup: The Future of Relational Foundation Models, Systems, and Real-World Applications
For Principal ML Engineers
Your custom models are limited by flat feature tables that destroy relational signals. KumoRFM learns from the full relational graph, discovering features your team would never think to engineer. It's also pre-trained on thousands of relational datasets, adding pattern knowledge that delivers an additional 10-50% accuracy boost on RelBench. And it replaces months of pipeline work with hours. Zero feature stores, zero custom ETL, zero serving infrastructure.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Why Principal MLEs choose Kumo
Zero
Feature stores to maintain
KumoRFM discovers features directly from your relational schema. No feature store, no feature pipelines, no train-serve skew. The foundation model handles feature engineering.
One API
For 50+ prediction types
Churn, fraud, recommendations, LTV, demand — all served from the same platform. One integration to maintain instead of N custom serving stacks.
Continuous
Retraining built in
Models automatically retrain as new data arrives. No cron jobs, no data drift monitoring hacks, no manual retraining pipelines to maintain.
Loved by data scientists, ML engineers & CXOs at
In production today
Enterprise Customer
Went from 3 models in production to over 50 in a single quarter with the same team. Feature engineering eliminated, pipeline complexity reduced by 95%.
Databricks
Lead-scoring models delivering dramatic improvement in conversion rates. Deployed in days instead of months, with zero feature engineering.
DoorDash
Restaurant recommendations driving hundreds of millions in GMV. Expanded to notification reranking and send-time optimization using the same foundation.
The challenge you know too well
Path 1 — Custom pipelines per model: Every model needs custom ETL, feature computation, serving, monitoring, and retraining — and you're responsible for all of it. The pipeline code dwarfs the model code, and every new use case multiplies the maintenance burden.
Path 2 — LLMs for structured data: They need entirely new infrastructure — GPU clusters, prompt engineering, guardrails, latency optimization — for problems your data warehouse already has the answers to. And they still can't reason over relational structure.
Kumo connects directly to your data warehouse and handles the entire lifecycle — training, deployment, monitoring, and retraining. No feature stores. No custom ETL. Define predictions in a query language and the foundation model handles everything else.
95%
Less data preparation
Feature engineering eliminated
10–50%
Accuracy improvement
Over traditional ML (RelBench)
20x
Faster time-to-value
From months per model to hours
55+
Use cases
Validated in production
KumoRFM isn't limited to your data alone. Pre-trained on billions of relational patterns across diverse datasets and fine-tuned to your schema, it sees what no in-house model can. As per the SAP SALT benchmark.
LLM
GPT4 + AutoML
PhD Data Scientist
Feature eng. + XGBoost
KumoRFM
Relational Foundation Model
17x
increase in models shipped per quarter
“Beating internal XGBoost model on key metrics with far less data and features. We went from three models in production to over fifty in a single quarter, with the same team.”
Matt Loskamp
GTM Data Science Leader, Enterprise Customer
From startups to enterprises, leading organizations rely on Kumo to deliver predictive insights at scale.
Peer-reviewed
Kumo is built on 40+ peer-reviewed papers at NeurIPS, ICML, and KDD. The methodology is public, reproducible, and benchmarked against the best methods in the field.
K. Huang, M. Fey, J. Leskovec et al.
A foundation model for relational data - pre-trained across schemas, it delivers accurate predictions out of the box and improves with fine-tuning on your specific data.
Read paperM. Fey, W. Hu, K. Huang, J. Leskovec et al.
Introduces learning predictive models directly on relational databases, eliminating the feature engineering pipeline that has historically bottlenecked enterprise ML.
Read paperJ. Robinson, R. Miao, K. Huang et al.
An open benchmark for evaluating relational prediction methods across 11 databases and 30 tasks. Kumo consistently outperforms traditional ML baselines.
Read paper
