Trustworthy AI requires a grounded foundation.
Start making AI show its work on your data. Intelligible makes the evidence visible, reusable, and available to any AI.
CSV is on every plan. Databases and Fivetran sources are on Enterprise.
Chat with your data. Upload a CSV, or connect a database on Enterprise, and ask questions. Summand answers from the models it fitted and the statistics it computed, and agents can call the same results over MCP.
See it in action on a demo dataset →Elbi is open source and runs on your machine. Pull in your sources, write derivations in Python or wrap a model from MLflow, and serve them to agents over MCP as cached tools. Its chat agent attaches a verification verdict to every answer.
uv tool install elbiTry it now →Why people pick Intelligible
- Results are versioned
- Summand writes a new version every time a component re-runs, and the old versions stay where they were.
- You can read the model
- Summand's feature effects come from an Explainable Boosting Machine. EBMs are additive: an analyst can read off what the model learned about each column, and about each pair of columns that interact. EBMs came out of Microsoft Research, in work led by Rich Caruana, who is now our Chief Science Officer.
- Elbi's checks are open source
- Elbi's verification gates, SDK and spec are Apache-2.0. The gates run in your own Python process, and each one's source is in the repo.
Questions we get
What's the difference between an empirical layer and a semantic layer?
A semantic layer holds what people wrote down, like metric formulas and business rules. It's standardised now as Apache Ossie (formerly Open Semantic Interchange, or OSI), and Elbi imports and exports it. An empirical layer, which Summand and Elbi both build, holds what the data does: distributions and effects computed against the actual rows.
Does Summand train on our data?
No. Summand fits models on your data for your workspace only. Its assistant runs on Anthropic's Claude, and Anthropic's API receives what a question needs, such as computed statistics, column descriptions and a few sample category values, plus the rows a query returns (at most 1,000). Workspace search indexes dataset metadata with Cohere's embedding model on AWS Bedrock.
Where does our data end up?
Summand's connectors copy your source on a schedule into Iceberg tables in Summand's AWS account (us-east-1). Fivetran handles the managed-pipeline sources and writes into the same account. Summand never writes back to your warehouse or deploys into your cloud. Elbi runs wherever you install it.
What happens when the data changes?
Scheduled components re-run and write a new version. Old versions stay as they were. For feature effects and column stats you can overlay two versions to see what moved.
Is there a free tier?
Elbi is free and open source. Summand's free tier takes CSV uploads up to 50 MB, keeps three datasets active at once and answers 20 questions a week; database connectors start on Enterprise. The pricing page has the rest.