AI · Advanced
Advanced AI Practice
Agents, RAG, evals, structured output, cost and guardrails.
What you will learn
The 8 topics inside Advanced AI Practice
Each topic has its own goal, its own questions and its own mastery score, so you always know which part is still weak.
Agents and tool use
Decide when an agent loop beats a single call.
Cost and latency
Reason about what a design costs per request.
Evaluation
Measure quality instead of eyeballing it.
Fine-tuning and customisation
Decide when tuning beats prompting or retrieval.
Guardrails
Contain the damage a wrong or hostile output can do.
Multi-agent orchestration
Design coordination between multiple cooperating agents.
Retrieval pipelines
Trace a question through a RAG pipeline and find where it fails.
Structured output
Get output a program can consume safely.
Example questions
Real questions from this course
These are taken straight from the course. Pick an answer, then open the explanation to see why it is right.
Sample 01
What makes a tool definition good?
Show the answer and why
Answer: The definition should include a precise name, clear usage instructions, and a strict parameter schema for the model.
A precise name, a plain description of when to use it, a tight typed parameter schema, and errors returned as readable text the model can act on.
A precise name, a plain description of when to use it, a tight typed parameter schema, and errors returned as readable text the model can act on.
Sample 02
What system performance impact must developers account for when designing centralized orchestrator architectures for real-time applications?
Show the answer and why
Answer: Significant orchestration latency overhead from recurring model calls and state serialization.
Routing requests through multiple language model calls adds network and token processing delays, defending against the misconception that orchestration overhead is negligible.
Centralized routing and agent-to-agent communication introduce significant orchestration latency overhead through repeated model calls and state transfers. Developers must evaluate whether multi-agent delegation justifies the added delay over a single-step agent.
Sample 03
In strict schema enforcement mechanisms, how should optional fields be modeled in JSON Schema if all fields must technically be marked as required?
Show the answer and why
Answer: By marking optional fields as nullable using union types while keeping them in the required list.
Strict schemas require all properties to be listed in the required array. To allow missing or null values, developers use union types like type [string, null]. The misconception defended against is assuming optional fields can simply be omitted from the required list in strict mode.
Optional fields should be represented using union types that allow the value to be either the target type or null, while remaining in the required array.
Advanced AI Practice is free with an account, along with every other course in the library.
Start free