v1.0.0 — April 2026
Initial Release
- Model registry: Central catalog for all ML models across the organization
- Model metadata: name, description, use case, risk tier, owner team, framework, tags
- Risk tiers: Low, Medium, High, Critical
- Data sensitivity levels: Public, Internal, Confidential, PII, Restricted
- Model statuses: Draft, Active, Deprecated, Archived
- Version management: Full model version lifecycle with approval workflows
- Stages: Development, Staging, Production, Retired
- Training metadata: dataset, sample count, hyperparameters, performance metrics
- Artifact storage with URI-based references
- Promotion and approval workflows with audit trail
- Model monitoring: Real-time prediction monitoring with configurable metrics
- Built-in metrics: latency, throughput, error rate, prediction distribution
- Custom business metrics with user-defined aggregation
- Configurable alert rules with threshold, rate-of-change, and anomaly triggers
- Alert channels: webhook, email, Slack
- Drift detection: Statistical drift detection for features and predictions
- Statistical tests: Kolmogorov-Smirnov (KS), Population Stability Index (PSI), Chi-Square, Jensen-Shannon Divergence
- Drift types: Data drift, concept drift, prediction drift
- Configurable thresholds per feature and per model
- Automated scheduled evaluations with alerting
- Fairness evaluation: Bias audits across protected demographic attributes
- Metrics: Disparate Impact, Equalized Odds, Demographic Parity, Calibration, Predictive Parity
- Protected attributes: gender, ethnicity, age group, geography, income band
- Threshold-based pass/fail with configurable acceptable ranges
- Historical evaluation tracking for trend analysis
- Explainability: Model explanation generation for individual and batch predictions
- Methods: SHAP values, LIME, feature importance, partial dependence plots
- Human-readable explanation reports for regulators and auditors
- Global and local feature importance
- A/B experiments: Controlled model comparison experiments
- Configurable traffic splits with gradual rollout
- Statistical significance testing (frequentist and Bayesian)
- Guardrail metrics with automatic experiment stopping
- Experiment statuses: Draft, Running, Paused, Completed, Cancelled
- Multi-tenancy: Full tenant isolation for all resources
- Sandbox: Full sandbox environment with synthetic models and prediction data