Topic 7
Module wrap-up
This lesson ends the module, and the course's main sequence. Production AI systems are mostly ordinary software engineering around a component that's probabilistic, changes underneath you, costs money per token, and can say anything.
Scroll sideways to see the whole diagram
Which model for which request, behind a gateway.
Text version
- Which model for which request, behind a gateway.
- What happens when a model fails.
- What it may cost, how fast, and for whom.
- How you know it's working, and what to do when it isn't.
- What it may say, and how changes reach users.
- When changing the model itself is worth it, and the checklist that ties the module together.
Predict first
Module wrap-up. Across the module:
- Choosing and routing models: the best model for your application, pinned versions, routers, cascades and a gateway.
- Fallbacks and graceful degradation: independent fallbacks, circuit breakers, degraded modes and load shedding.
- Budgets and usage controls: cost models, safe caching, rate limits, quotas and latency budgets.
- Observability and incidents: metrics from failures, breakdowns, traces, drift, feedback and postmortems.
- Guardrails and safe deployment: input and output guardrails, flags, shadow and canary releases.
- When to fine-tune: form against facts, the ladder of adaptations, the real costs, and a readiness plan.
From this lesson, you can now:
- place fine-tuning after prompting and retrieval, and say why;
- name good reasons to fine-tune, and the costs that argue against it;
- classify failures as information or behaviour, and pick RAG or fine-tuning accordingly;
- estimate fine-tuning memory, and explain what parameter-efficient methods like LoRA save;
- write a production-readiness plan as a ranked list of gaps.
Quick check