What problem does it solve? Moving a trained model from a notebook to a production endpoint involves serving frameworks, containerization, scaling, and rollout strategy decisions that are easy to get wrong. This Skill provides tested patterns and runnable scripts for exposing models as real-time or batch inference services. ## Core Features & Use Cases - API Serving: Build FastAPI model servers with health checks, Prometheus metrics, input validation, prediction caching, and A/B traffic splitting, or use BentoML for packaged services. - Batch Inference: Run large-scale scoring over CSV, Parquet, or JSON data with sklearn, ONNX, or PyTorch models, including per-batch latency stats and failure handling. - Production Deployment: Containerize with Docker, deploy to Kubernetes with HPA autoscaling, and roll out via canary, blue-green, shadow, or A/B patterns; optimize inference with ONNX Runtime, quantization, and TensorRT. - Use Case: After training a fraud-detection model, serve it directly from the MLflow registry with serve_model.py --model-uri "models:/fraud-detector@champion", then roll out v2 to 5% of traffic using the built-in A/B ratio flag. ## Quick Start Ask the agent to serve your trained model file as a FastAPI endpoint with health checks and metrics using the serve_model.py script.
