Epiphany

From pipeline output to trained model

Describe the data work and the modeling goal. Epiphany deploys the pipeline, opens ML Lab on the output, trains a baseline, and attaches the model — no separate ML stack.

<1hr
Pipeline to model
2min
Budgeted search
100%
Computed metrics

How It Works

01

Open from the pipeline

One click from a pipeline or a Data Lab session. Agents reuse profiled columns — no export, no notebook bootstrap.

02

Template or scan → train

Confirm a target from the scan, pick a template, train a baseline. Iterate with agent suggestions on the same session.

03

Evaluate and attach

Metrics stay on the experiment. Deploy attaches the model and retrain job to the source pipeline.

Key Benefits

Describe the modeling goal in the same sentence as the pipeline
One-click from pipeline output or a Data Lab session into ML Lab
Pre-built templates: classification, regression, forecasting, anomaly
Agents scan targets, train a baseline, then iterate experiments
Evaluation and artifacts stay next to the source pipeline
Deploy the model back onto the same scheduled job

What You Get

Pipeline Integration

Model training attaches directly to the source pipeline. Score step writes predictions back to the warehouse on every run.

Budgeted Search

Optional 2-minute model search across HGB, linear models, and LightGBM/XGBoost. Leaderboard with timing.

Versioned Experiments

Every train creates a versioned experiment (v1, v2, ...) with computed metrics, confusion matrix, and feature importance.

Model Artifacts

Trained models stored as joblib in Cloudflare R2. Online scoring API, rollback support, presigned download URLs.

Real-World Examples

Customer Churn Prediction

Classify which customers are likely to cancel based on transaction history and usage patterns

Sales Forecasting

Predict future revenue from historical sales data with lag features and seasonality handling

Fraud Detection

Flag anomalous transactions using isolation forest and statistical outlier detection

Lead Scoring

Rank prospects by conversion probability using CRM data and engagement signals

ML Templates

Each template sets up the right task type, split strategy, and metrics

Classification (Binary & Multi-class)
Regression
Time Series Forecasting
Anomaly Detection

Built On

The ML stack under the hood

Scikit-learn
XGBoost
LightGBM

Questions about this use case

Train writes a versioned joblib to Cloudflare R2. Promote attaches a score step to the source pipeline. The next job run downloads that artifact and writes {target}_predicted, {target}_proba, and model_version onto the outgoing table. Rollback swaps or removes that step.

Ready to train your first model?

Start building ML pipelines with automated infrastructure and deployment