Automated data science you can audit.
Summand is an AI-native integrated data science platform. Ask a question and it answers by running something: a column statistic, a SQL query, or a model it fits. It has twenty tools and chooses which to use, and training a model is one of them.
What the chat can do
Behind the chat are statistics, SQL, charts, views, schema comparison, model training and prediction, and counterfactuals. SQL runs on Athena, and the assistant reads back at most 1,000 rows per query.
Free accounts get 20 questions a week; Pro and Education are uncapped.
Try asking
- List my datasets
- Show me my feature importance graph
- Train a model that predicts churn
- What would have to change for this customer to stay?
- What did the last run flag as unusual?
Results live in the catalog
The catalog lists your connectors, their datasets, and the components computed on each one. Open a component and you get its chart and a version switcher.
Feature effects come from an Explainable Boosting Machine. EBMs are additive, and the shape function you see is the model itself. When Summand shows how square footage bends against price, that curve is what the model used to predict.
What you can open
- Feature effects per column, and pairwise interactions
- Column stats
- Any earlier version of a component
- Two versions overlaid on one chart
- A lineage graph of what feeds what
Experiments run on a schedule
An experiment is a set of components on a dataset, with a schedule. Components run in dependency order when data lands and again on that schedule, and each run writes a new version. A freshness checker watches the connectors. Summand flags schema drift on ingest, and drift in a model’s metrics.
20
chat tools
11
components to pick from
5
model algorithms
300+
sources via Fivetran
SQL and views
Write SQL in the editor and save it as a view. You can also build a view with the visual join builder, or have Summand draft one. Views can join across connectors, because every source lands in the same store.
Models
Train an EBM, CatBoost, LightGBM, XGBoost or a linear model. Hyperparameters are tuned automatically, and metrics are reported on a 20% holdout. Pick one of the tree models when accuracy matters more than a curve you can read. Models are versioned (compare two, promote one, archive the rest).
| Algorithms | EBM, CatBoost, LightGBM, XGBoost, linear |
| Tuning | Automatic, time-boxed |
| Metrics | Accuracy, log loss, RMSE, MAE, R² on a 20% holdout |
| Lifecycle | Versioned, comparable, promote or archive |
Where your data comes from
Summand connects directly to Postgres, MySQL, Snowflake, Azure SQL and Delta Sharing, and reaches everything else through Fivetran.
CSV is on every plan. Databases and Fivetran sources are on Enterprise.
Agents connect over MCP
Claude, ChatGPT or your own agent can reach the same datasets and components through Summand’s MCP server. The API page lists what they can call.
Docs
Read more about what makes Summand unique: connectors, views, experiment, and components.