---

name: mlops-patterns description: Follow these patterns when implementing MLOps features in OptAIC. Use for ML model definitions (5-component structure), model instances, training/inference pipelines, model registry, and monitoring. Covers signal models, macro regime models, relevance models, and signal combining/filtering models.

MLOps Implementation Patterns

Guide for implementing MLOps features that integrate with OptAIC's resource-based architecture.

When to Use

Apply when:

• Creating ML Model Definitions (MLModuleDef) with 5 code components
• Implementing Model Instances in MLOps Center
• Building training, inference, or monitoring pipelines
• Integrating with model registry (MLflow or internal)
• Implementing model categories (signal, regime, relevance, combining)

MLOps Three-Tier Model

CopyMLModuleDef (Definition)    ModelInstance (Config)       Execution (Runs)
────────────────────────    ──────────────────────       ─────────────────
XGBSignalModelDef       →   SPX_Alpha_Model          →   TrainingRun
  (5 code components)         (datasets + config)          InferenceRun
                                                           MonitoringRun
                                                               ↓
                                                         ModelVersion

ML Model Categories

Implementation Workflow

1. Create MLModuleDef (5 Components)

CopyMLModelDef/
├── model/           # Model architecture + hyperparameter schema
├── training/        # Trainer + evaluator
├── inference/       # Predictor + batch inference
├── monitoring/      # Data drift + performance monitoring
├── tests/           # Test suite for all components
└── docs/            # Documentation

See references/mlmodule-structure.md.

2. Create Model Instance

Compose MLModuleDef + datasets + config. See references/model-instance.md.

3. Implement Pipelines

• TrainingPipeline → reads datasets, produces ModelVersion
• InferencePipeline → reads features + model, writes predictions
• MonitoringPipeline → reads data/preds, emits metrics/alerts

See references/mlops-pipelines.md.

4. Integrate with Registry

See references/model-registry.md.

5. Create UI Components (MLOps Center)

Two views required:

• Model Instance View - registered models with configs
• Execution View - training, registry, inference, monitoring

See references/mlops-center-ui.md.

Critical Rules

1. 5-component structure - MLModuleDef must have model, training, inference, monitoring, tests
2. Activity emission - All runs emit activities (training, inference, monitoring)
3. Lineage tracking - Link dataset versions → model version → prediction dataset
4. Guardrails - Validate model outputs (e.g., signal bounds)
5. PIT correctness - No lookahead in training or inference

Tech Stack

Unified ML SDK (optaic.mlops)

All MLOps infrastructure is wrapped in a unified SDK for seamless development:

Copyfrom optaic.mlops import tracking, registry, monitoring, pipeline
from optaic.mlops.base import BaseModel, BaseTrainer
from optaic.mlops.data import load_dataset

Key modules:

• tracking - Experiment logging (wraps MLflow)
• registry - Model versioning (wraps MLflow Model Registry)
• monitoring - Drift & performance (wraps Evidently)
• pipeline - Orchestration (wraps Prefect)
• data - PIT-aware dataset access
• base - Base classes for model definitions

See references/unified-sdk.md and Blueprint section 8.9.

Reference Files

• Unified SDK - optaic.mlops SDK patterns
• MLModuleDef Structure - 5-component package
• Model Instance - Configuration patterns
• MLOps Pipelines - Training/inference/monitoring
• Model Registry - Version management
• MLOps Center UI - Two-view architecture
• MLflow + Evidently Integration - Experiment tracking & monitoring