Course outline for Applied Generative AI

Module 1: Specify and validate an interaction (2 hours)

• Prompting: Task instructions, examples, exclusions, and output requirements; separate model inference from calculations that deterministic code should perform.
• Structured outputs: Required fields, allowed values, and schema checks; validate source support and intended meaning separately from structural compliance.
• Tool use: Capability descriptions, proposed arguments, and external execution; treat model-selected operations as untrusted proposals rather than granted authority.

Practical outcome: Convert a document classification request into a prompt, schema, and semantic rubric; distinguish model inference from deterministic calculation and treat proposed tool arguments as untrusted.

Module 2: Choose a bounded execution path (2 hours)

• Function calling: Argument schemas, call identifiers, and tool-result exchanges; reject malformed inputs before invoking the supplied mock function.
• Model selection: Task criteria, held-out response comparisons, and observed error categories; choose models using evidence rather than brand or self-reported confidence.
• Model routing: Simple and difficult case routes, error costs, and review limits; compare fixed routing policies against observed failures and abstentions.

Practical outcome: Inspect a mock function-call exchange, compare two model response sets on held-out inputs, and choose routes using false-positive, false-negative, and human-review costs rather than brand or stated confidence.

Module 3: Use context and handle uncertainty (2 hours)

• Context engineering: Relevant evidence, examples, and minimal task inputs; exclude unnecessary data and distinguish missing support from an unclear instruction.
• Multimodal interaction: Text and image evidence, input relevance, and interpretation limits; construct bounded requests without treating uncertain perception as established fact.
• Decision and classification models: Label contracts, false positives, false negatives, and abstention thresholds; select policies on development cases before evaluating held-out results.

Practical outcome: Prepare minimal text/image context, include an abstention route for insufficient evidence, and compare threshold policies against observed errors and a fixed reviewer capacity.

Practical exercise

• Produce a classification prompt, output contract, routing table, and held-out evaluation report.
• Accept a valid source-supported classification; reject a structurally valid but unsupported label and malformed function arguments.
• Score at least twelve supplied cases, separating errors, abstentions, and review demand; compare two routing policies against a stated review limit.
• Explain one case where rules or human review are preferable and why self-reported confidence is not a calibrated probability.