Automated machine learning · invite-only beta
From raw table to a model you can defend.
Profile the data, review the reasoning, run the search and keep the pipeline. AutoML makes every choice inspectable before your model ships.
Every rowprofiled, never sampled
Every decisionreasoned and editable
Every artifactexportable and yours
Product tour · 00:24
See the entire run, not a staged mockup.
Real footage of the bundled housing dataset moving through profiling, schema lock, configuration, live training and results.
The workflow
Five moves from a raw file to a served model.
Each stage pauses for your approval and leaves a record your team can inspect later.
-
01
Connect data
Drop in CSV or Parquet. AutoML checks, counts and profiles every row before the run begins.
-
02
Lock the schema
Review inferred types, nullability and the target. Correct the details, then lock a version.
-
03
Review the plan
Get a drafted job configuration with a written reason behind every field and split.
-
04
Run the search
Watch five progressive passes test baselines, boosted trees, interactions and ensembles.
-
05
Ship the winner
Export the pipeline, model artifact and audit trail, or serve the best run behind an endpoint.
Human in the loop
Nothing happens that you can’t inspect.
Every inferred type, dropped column, split strategy and hyperparameter is recorded with the evidence behind it. Change any of it before compute starts.
- 01 Leakage-aware validation
- 02 A reason for every decision
- 03 Versioned, readable Python
- 04 Full run history
287 search = CandidateSearch(
288 families=["lightgbm", "xgboost"],
289 metric="roc_auc",
290 validation=TimeSeriesSplit(n_splits=5),
291 budget=Budget(trials=40),
292 )
293
294 best = search.fit(X, y).leader()
295 best.export("artifacts/housing_r5")
Bring your own compute
Train where your data already lives.
Link a laptop, lab box or cloud instance. The console orchestrates while training and sensitive rows stay on the machine you chose.
Your table is enough to start