MDC | AI-Powered Data Tools

Choose a tool to get started

Mass Data Compare

Migrate data with confidence

Cross-database validation across multiple clouds/databases. Schema, count, PK distribution & hash comparison strategies.

Data Drift Detection

Actively listen to your data changes

Database-connected drift detection with KL divergence, PSI, equal-frequency binning. Compare baseline and current distributions across selected columns and time windows.

Data Explorer

Fast schema summary and lazy column profiling on CSV or warehouse tables. Distributions, quality flags, and numeric correlations from a stable sample.

● Data Source

● Table & sample

Choose the table to profile and the stable sample size for lazy column stats.

Table name

Sample size

Columns

Schema from summary. Load distributions per column or in batches.

Column detail

Quality flags

Correlations

▦ Raw Data Samples

Expand

Inspect sample rows from the connected table using the same column names as in the schema.

Columns

Rows

Run EDA, then refresh samples to inspect raw rows.

⚖ Model Identity

Training Date ?Training DateWhen the model was last trained. Used to assess data staleness and concept drift risk. Example: a model trained in 2019 misses COVID-era shifts.

Domain ?DomainThe industry your model operates in. High-sensitivity domains like lending and finance get a higher concept drift multiplier because market conditions change faster.

⚙ Model Properties

Target Variable ?Target VariableThe column your model predicts. Example: loan_status (default vs. paid), fraud_flag (fraud vs. legitimate).

Features (comma-separated) ?FeaturesThe input columns used by the model. Checked for data leakage — features that contain future information not available at prediction time. Example: recoveries in a loan default model is leaky because recovery only happens after default.

Class Distribution ?Class DistributionHow the target variable is split between outcomes. Used to detect class imbalance. Example: if 97% of loans are "Fully Paid" and only 3% "Charged Off", the model may just predict the majority class and still look accurate.

Class Name (Majority)

Count / Proportion

Class Name (Minority)

Count / Proportion

Reported Metrics ?Reported MetricsPerformance numbers from model training/evaluation. Cross-referenced with class distribution to flag suspicious results. Example: 99% accuracy with severe imbalance likely means the model just predicts the majority class.

Accuracy ?Accuracy% of all predictions that were correct. Can be misleading with imbalanced classes.

AUC / ROC ?AUC / ROCArea Under the ROC Curve. 1.0 = perfect, 0.5 = random. Values > 0.99 are flagged as suspicious.

F1 Score ?F1 ScoreHarmonic mean of precision and recall. A big gap between accuracy and F1 signals poor minority-class performance.

Recall (Minority) ?Recall (Minority)Of all actual positive cases, what % did the model catch? 50% recall means half go undetected.

$ Financial & Risk

Annual Revenue ($) ?Annual RevenueYearly revenue influenced by the model's predictions. Used to estimate expected annual loss from a stale model. Example: $6M revenue with 15% risk exposure = $900K potential loss.

Refresh / Retrain Cost ($) ?Retrain CostHow much it costs to retrain the model (compute, data engineering, validation). If expected annual loss exceeds this cost, the advisor recommends retraining. Example: $30K retrain cost vs. $135K expected loss → ROI of 4.5x.

Revenue Impact Factor ?Revenue Impact FactorFraction of revenue at risk due to model degradation. Example: 0.15 means 15% of revenue could be impacted by a stale model.

⚙ Advanced Tuning (Optional)

For PSI/KL drift against live data, use Drift Monitor. Paste ADWIN time-series JSON here to enhance concept-drift scoring.

Load ADWIN Sample Scenario

ADWIN Drift Data — Time-Series JSON Array

Correlation Data (JSON) ?Correlation DataJSON object mapping feature pairs to correlation coefficients. Used to detect multicollinearity and feature redundancy.

Custom Weights (JSON) ?Custom WeightsOverride the default check weights used in scoring. JSON object mapping check names to weights (0-1). Example: {"staleness": 0.3, "feature_leakage": 0.2}

Demo workspace

Pre-loaded test data. Users cannot create workspace.

☷ Drift Detection Configuration

● Baseline (Training)

● Current (Detection)

Same as baseline (training)

Model ID

Table Name (shared for database; CSV uses each card’s file)

Date Column

Sample Size (default: 20,000 rows)

Metric Precision

Baseline Window

From

Detection Window

From

Statistical Tests

Columns to Monitor

Available 0

Selected 0

Upload a CSV or browse a table to populate columns

▦ Raw Data Samples

Expand

Inspect familiar raw rows from the active dataset and compare them with the drift metrics above.

Dataset

Window Preset

Rows

Load a dataset and click Refresh samples to inspect raw rows.

Rolling Drift Analysis

Expand

Split your data by a date column into time windows and track how drift evolves over time. Rolling analysis uses the current dataset. The first window in the selected rolling range is used as the baseline; each subsequent window is compared against it.

Rolling From

Rolling To

Frequency

Ccube Data Products

Mass Data Compare

Data Drift Detection

Data Explorer

Comparison Final Result