ml-engineer_skill

Overview

This skill implements production-ready ML pipelines, model serving, and robust feature engineering. It focuses on reliable deployment of TensorFlow, PyTorch, and ONNX models with A/B testing, monitoring, and drift detection. Emphasis is on production reliability, latency targets, and observability rather than pushing complex research prototypes.

How this skill works

I build end-to-end pipelines that start with a simple baseline model and iterate toward production requirements. The skill versions data, features, and model artifacts, creates scalable serving endpoints (TorchServe/TF-Serving/ONNX), and adds A/B testing and gradual rollouts. It also installs monitoring, drift detection, and alerting with retraining hooks for automated lifecycle management.

When to use it

• Deploying TF/PyTorch models to production with strict latency or throughput SLAs
• Setting up feature pipelines with validation and version control
• Running A/B tests or canary rollouts for new model versions
• Implementing batch or real-time inference at scale
• Adding monitoring, drift detection, and automated retraining triggers

Best practices

• Start with a simple baseline and incrementally add complexity once production constraints are clear
• Version everything: raw data, feature transforms, model artifacts, and serving configs
• Define latency and throughput SLOs early and optimize inference paths (quantization, batching, ONNX)
• Implement gradual rollouts and automated rollback procedures for safe deployments
• Instrument prediction pipelines with quality metrics, logging, and alerts for drift

Example use cases

• Serve a PyTorch image classifier via TorchServe with autoscaling and warmup to meet 100ms p95 latency
• Create a streaming feature pipeline that validates inputs, materializes features, and records lineage
• Run A/B experiments comparing two model versions with traffic split and offline metrics aggregation
• Deploy an ONNX optimized model for cost-efficient CPU inference with batching and latency constraints
• Set up monitoring that detects label or feature drift and triggers retraining workflows