Module 25
MLOps and LLMOps in Production
A model that worked on launch day quietly gets worse as the world changes, and an LLM app can break from a one-line prompt edit or a surprise bill. This module follows Crumb’s croissant forecaster and WhatsApp bot through the full production lifecycle: deploying, monitoring and drift, versioned prompts behind eval gates, cost control, and the governance records a real customer will ask to see.
By the end of this module you can
- explain the MLOps loop and tell data drift from concept drift, with a working drift check
- run an LLM app with versioned prompts, per-category eval gates, canary releases and feedback loops
- measure cost per request and per conversation, then cut it with compression, caching and routing
- write a model card, an append-only audit log and an enforced data-retention rule
- choose between prompting, RAG and fine-tuning with a clear decision table
Lessons
4 in this module- 01The MLOps lifecycle: deploy, monitor, driftYou can walk through the MLOps loop, tell data drift from concept drift, and build a simple weekly monitor that checks data quality, input drift and prediction error.
- 02The LLMOps lifecycleYou can run an LLM feature through ideation, development and operations, with fingerprinted prompt versions on every log line, a per-category eval gate, canary releases and a feedback loop that turns failures into tests.
- 03Cost management and prompt compressionYou can measure the cost of every call and conversation, find where the tokens go, and cut the bill with prompt compression, prompt caching, model routing and token budgets while evals guard quality.
- 04Governance, readiness and RAG vs fine-tuningYou can back every governance claim with evidence (a model card, a tamper-evident audit log, enforced retention and PII redaction, recorded approvals), run a production-readiness checklist, and choose between prompting, RAG and fine-tuning.