LLMOps

The operations surface for an LLM-backed product: deployment patterns, model promotion gates, prompt versioning, traffic shaping (shadow / canary / A-B), and the eval-driven release loop.

Interview talking points

• How do you ship a prompt change safely? Rehearse the pre-merge eval gate (golden set + LLM judge), then the canary plan with auto-rollback on acceptance-rate drop.
• Prompt vs model vs retrieval — which do you tune first? The order is retrieval (highest signal/effort) → prompt → model swap → fine-tune. Why each in that order is in here.
• What's in your release pipeline? Map this onto the CI/CD User Stories — cross-reference questions are common.
• How is LLMOps different from MLOps? Prompt as artefact, eval as judgement, no gradient updates in the hot loop.

Files in this folder

Back to the home page.