ML Serving MLOps Review

Use this skill for production ML serving, MLOps, deployment review, monitoring, retraining, model cards, release gates, and ML debugging triage.

Process

1. Identify the serving mode: batch, online, streaming, edge, embedded library, managed provider, or model server.
2. Inspect model contracts, input/output validation, feature manifests, preprocessing versions, artifact formats, fallback behavior, and train-serving parity tests.
3. For real project work, detect repo-pinned model/runtime/framework versions first. Use itsol-current-tech-context for current official docs before judging fast-moving model-serving APIs, registries, providers, or security defaults.
4. If Rust/Rig/Candle/provider runtime code is the primary surface, route first to rust-ml-llm-architecture, rust-ml-llm-debugging, or rust-ml-llm-review; use this skill for cross-cutting serving, MLOps, rollout, monitoring, and review concerns.
5. For large, multi-surface, production-impacting, security/privacy-sensitive, or release-sensitive ML PR reviews, use focused subagents before the final verdict. Split review by risk area: data/evaluation, training/experiments, LLM/RAG, serving/MLOps, security/privacy, QA/release, and rust-ml-llm-review when Rust/Rig/Candle code is primary.
6. Lead review output with concrete findings, severity, affected behavior, missing verification, rollback risk, and file references.

Evidence

Prefer model contracts, schema tests, parity tests, registry records, model cards, CI logs, training run metadata, deployment manifests, monitoring dashboards, incident logs, eval reports, and release approvals over assumptions.

Focused References