FlowRunner
PricingContact
Theme
Start Free

BigML

AI

Train and run machine learning models with BigML, covering sources, datasets, models, predictions, batch scoring, and evaluation. Agents score records against a trained model as a step inside a business workflow.

22 actions API key available
BigML website ↗ Platform Documentation ↗ Capability data verified 2026-08-01
A monthly retrain fires on the latest closed and churned account data
Agent calls Create Source From URL on the export, then polls Get Resource until the status reports finished
Agent calls Create Dataset with the objective field pinned, then Create Ensemble to train on it
Agent calls Create Evaluation against a held-out dataset the new model never saw
Agent compares accuracy, precision, recall and the confusion matrix against the model currently in production
Agent posts both scorecards side by side with the classes whose recall moved
The analytics owner approves promotion, or leaves the current model in place and keeps the new one parked

What This Integration Enables

Every other service in this category answers a question by prompting a model. BigML does not. It trains one on your data and then runs it, which is a different trust proposition altogether. A decision tree is white box: you can open it and see the splits that produced a prediction. Create Prediction returns not just a value but a confidence and the per-class probabilities behind it. Create Evaluation returns accuracy, precision, recall, f-measure and a confusion matrix on a held-out dataset. None of that is available from a prompt, and all of it is what a risk owner or an auditor asks for when a score starts affecting who gets credit, who gets flagged, or who gets called.

The catch is that classical machine learning has always lived one step away from the process it should be informing. A model sits in a notebook, and the operational system that needs it has no way to call it. That gap is exactly what Orchestration as a Service closes: the score becomes a step in the workflow that created the record, running against a trained model rather than a rule somebody wrote three years ago. Because BigML is asynchronous, almost every job here is a submit-and-poll pattern, and Get Resource is the single action that reports progress for all of them. FlowRunner handles the waiting, and applies human-in-the-loop where the outcome deserves it.

Without FlowRunner

A score with no provenance A number arrives, and nobody can say which fields produced it
Models promoted on vibes The retrained version goes live because it is newer, not because it tested better
Scoring stuck in a notebook The model exists, but nothing in the business process can call it

With FlowRunner

Inspectable structure A decision tree can be read, and an ensemble reports confidence and per-class probability
Promotion gated on evidence No model reaches production until its evaluation has been compared to the one it would replace
Scoring inside the process A record is scored as part of the workflow that created it, not in a batch job somebody remembers to run

Use Case Scenarios

Scoring a record the moment it is created

A new account appears in the CRM. The flow calls Create Prediction against the current churn ensemble with the account's fields, and gets back a predicted class, a confidence value and the probability of each outcome. Accounts above the risk threshold post into the customer success channel in Slack with the probability attached, so the owner can see the difference between a 91 percent call and a 54 percent one. The score is a step in the account creation workflow, not a report nobody opens.

Batch scoring a book of business overnight

A monthly review needs every open account scored, which would be tens of thousands of individual calls. The agent uses Create Batch Prediction against the dataset instead, polls Get Resource until the job reports finished, and calls Download Batch Prediction to pull the results back as CSV. The rows land in BigQuery next to the previous run, so the analyst reviews movement between months rather than a single snapshot.

Anomaly scoring transactions as they arrive

Rather than writing threshold rules that go stale, the team trains an anomaly detector on historical transactions with Create Anomaly Detector, which learns what normal looks like without anyone labelling anything. Each new transaction is scored with Create Anomaly Score, returning a value between 0 and 1. Rows scoring near 1 are written to an exceptions table in Airtable for a person to work through. The rest post without interruption, and the queue is short enough that somebody actually clears it.

Human-in-Loop Highlight

The moment that matters on this connector is promotion. Retraining on fresh data is routine and should be automatic. Deciding that the retrained model replaces the one currently making decisions is neither. A new model can post a better headline accuracy while its recall on the one class you actually care about has quietly collapsed, and the confusion matrix is where that shows up. So the agent trains, evaluates against a held-out dataset with Create Evaluation, and then stops. It posts both scorecards to the analytics owner: "Retrained ensemble: accuracy 0.91 against the current model's 0.88. Recall on the churn class fell from 0.74 to 0.61, meaning it would miss more at-risk accounts than the model it replaces. Promote, retrain with a different sample rate, or keep the current model?" A person reads the matrix and decides. There is a second irreversibility worth naming on the same connector: Delete Resource cannot be undone, and deleting a project deletes every source, dataset, model and prediction inside it. That action belongs behind the same gate, for the same reason.

Agent processes routinely
Detects exception requiring judgment
Clear match Continues automatically
Ambiguous Routes to human via preferred channel
Human decides
Agent resumes with decision

Agent Capabilities

22 actions

Sources

3
  • Create Source From URL Creates a source by having BigML download a data file from a URL, accepting HTTP and HTTPS addresses as well as Amazon S3, Google Drive and Google Cloud Storage locations. Creation is asynchronous, so poll Get Resource until the source reports finished before building a dataset from it.
  • Create Source From Inline Data Creates a source directly from a string of data in the request, with no file hosting involved. Accepts CSV with a header row or a stringified JSON array. Used for small datasets produced by an upstream step in the same flow.
  • List Sources Lists the sources in the account, newest first, with pagination metadata and total count, filterable by name substring or tag. Used to resolve the source ID a dataset needs.

Datasets

2
  • Create Dataset Creates a dataset from a finished source, which is the structured, summarised form every model trains on. Fields can be narrowed with an input or excluded list, and a target column pinned as the objective field. Asynchronous, so poll Get Resource until it reports finished.
  • List Datasets Lists the datasets in the account with pagination metadata and total count, filterable by name substring or tag. Used to find the dataset ID required to train a model, cluster, anomaly detector or evaluation.

Supervised Learning

4
  • Create Decision Tree Model Trains a single white-box decision tree from a dataset, producing classification or regression predictions depending on whether the objective field is categorical or numeric. Used when the ability to read the model matters as much as its accuracy.
  • Create Ensemble Trains a collection of decision trees that vote on a combined prediction, which usually outperforms a single tree. Supports bagging, random decision forests and gradient tree boosting. Used as the default production model where accuracy leads.
  • Create Deepnet Trains a deep neural network for classification or regression, optionally letting BigML search for the best network structure automatically. Used for problems where tree-based models underfit, at the cost of longer training.
  • List Models Lists the decision tree models in the account with pagination metadata and total count, filterable by name substring or tag. Ensembles and deepnets have their own endpoints and are not returned here.

Predictions

4
  • Create Prediction Scores a single new instance against a finished supervised model and returns the predicted value with its confidence and per-class probabilities. Near real time, so the result usually arrives in the response. Used to score a record inside the workflow that created it.
  • Create Batch Prediction Scores every row of a dataset against a supervised model and writes the results to a downloadable CSV, optionally also to a new dataset. Used to score large volumes in one job rather than calling Create Prediction per row.
  • Download Batch Prediction Downloads the CSV produced by a finished batch prediction as text, ready to parse or forward to storage. The job must have reached finished status or the download comes back empty.
  • Create Evaluation Measures how well a supervised model performs on a held-out dataset, returning accuracy, precision, recall, f-measure and a confusion matrix for classification, or r-squared and mean error for regression. Used as the evidence a person reads before promoting a model.

Unsupervised Learning

4
  • Create Cluster Groups the rows of a dataset into clusters of similar instances with no labelled target column, either at a set number of clusters or letting BigML choose. Used for segmentation where the categories are not known in advance.
  • Create Centroid Assigns a single new instance to the nearest cluster of a finished cluster resource, returning the centroid name and the distance to it. Used to segment a new customer or record in real time.
  • Create Anomaly Detector Builds an isolation forest that learns what normal rows look like so unusual ones can be scored later, and reports the top anomalies already present in the training data. Used when the fraud or fault patterns are not known well enough to write rules for.
  • Create Anomaly Score Scores a single new instance against a finished anomaly detector and returns a value between 0 and 1, where values near 1 mark the row as strongly anomalous. Used to flag suspicious transactions, sensor readings or account activity as they arrive.

Projects and Resources

5
  • Create Project Creates a project, the container BigML uses to keep the resources of different initiatives apart. A source assigned to a project passes that project on to every dataset, model and prediction derived from it.
  • List Projects Lists the projects in the account with pagination metadata, total count and per-project statistics on how many resources of each type they hold. Used to navigate an account with several initiatives running.
  • Get Resource Retrieves any resource by its full ID and reports its progress, which is how every asynchronous job in this service is polled. The response carries a readable status label and a finished flag, so a flow can branch without decoding a status code.
  • Update Resource Updates the mutable fields of any resource with a partial object, most commonly to rename it or change its description, tags or category. Fields BigML does not mark as updatable are rejected.
  • Delete Resource Permanently deletes any resource by its full ID. Deletion cannot be undone, and deleting a project also deletes every resource inside it. Used for cleanup, and worth putting behind an approval for exactly that reason.

Frequently Asked Questions

What can FlowRunner do with BigML?

FlowRunner agents can run Create Source From URL, Create Source From Inline Data, and List Sources in BigML, plus 19 more actions.

Does connecting BigML to FlowRunner require OAuth?

No. BigML connects to FlowRunner with an API key, no OAuth flow required.

Can BigML trigger a FlowRunner workflow automatically?

BigML doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with BigML

$100 in credits. No card required. Connect in minutes.