edge-ai-overlay

name	edge-ai-overlay
description	Domain overlay for edge AI and ML deployment. Extends language standards with model lifecycle, inference optimization, and data pipeline rules.

Edge AI Overlay

Activates when sdlc-config.md sets domain.primary: edge-ai or domain.secondary includes it.

Additional ERROR Rules

• All model artifacts must be versioned with training metadata (dataset hash, hyperparameters, metrics)
• Model inputs must be validated for shape, dtype, and range before inference
• Never load models from untrusted sources without integrity verification (checksum/signature)
• Memory budget for inference must be documented and verified on target hardware
• All data pipelines must be reproducible — pin random seeds, document transforms, version datasets

Additional WARNING Rules

• Use ONNX, TFLite, or framework-native formats for deployment — not pickled Python objects
• Quantize models (INT8/FP16) for edge deployment — document accuracy impact
• Implement graceful degradation when model confidence is below threshold
• Pre/post-processing must match training pipeline exactly — test with golden samples
• Monitor inference latency and throughput in production — alert on drift
• Use batch inference where latency constraints allow

Additional RECOMMENDATION Rules

• Consider model distillation for resource-constrained targets
• Use TensorRT, OpenVINO, or Core ML for hardware-specific optimization
• Implement A/B testing infrastructure for model updates
• Consider federated learning for privacy-sensitive training
• Use feature stores for consistent feature computation between training and serving

MLOps Checklist

• Training data versioned and documented
• Model card with performance metrics, limitations, and bias analysis
• Inference benchmark on target hardware
• Rollback plan if model update degrades performance
• Monitoring for data drift and concept drift
• Audit trail for model decisions (explainability)