AI Apps, Agents & APIs

1. Calling a model from your server

Every AI feature in production is still a program: a request with inputs, a timeout, a failure mode, and a cost — wrapped around one call to a model that may not answer the way you hoped. This lesson covers where the call belongs (always behind your server), the request/response shape of a modern chat API, streaming and why it changes perceived latency, timeouts, retries, fallbacks and refusal handling, structured outputs, token accounting, and the caching that makes the economics work. The model is the easy part; the engineering around it is the job.

2. The prompt is product surface

When your product's behaviour is a prompt, the prompt is not a string in a file — it is the specification of what users experience, and it deserves the discipline the earlier courses gave to any other specification: versioned, reviewed, tested, and changed with evidence. This lesson covers system prompts as product constitution, structuring instructions for models, guardrails as code rather than prose, prompt-injection defences, few-shot examples as design, change management for prompts, and the tooling that makes prompt work legible. Users feel every word; treat the words accordingly.

3. Your own data: retrieval (RAG)

Models do not know your product, your documents or last week's decisions — and fine-tuning is rarely the right answer for facts that change. Retrieval-augmented generation fixes this by finding the few passages that matter and putting them in the prompt, so answers are grounded in your corpus with citations to prove it. This lesson covers chunking, embeddings and vector search, when retrieval helps and when it doesn't, hybrid search and reranking, building a citation-first pipeline, and the failure modes — stale indexes, retrieval misses, injected documents — with their fixes. Ground every answer, or label it as guessing.

1. Tools and agents

Giving a model tools — functions it can call — turns a text generator into a system that acts on the world, and every security lesson you have ever had becomes relevant at once: the model proposes, your code decides, and the permission checks live on your side of the boundary. This lesson covers tool design, the agent loop, permissions and confirmation for consequential actions, memory and planning, idempotency for model-driven writes, and the failure modes unique to autonomous action. The power and the risk are the same thing — design them together.

2. Evals, latency and cost

AI features fail softly: a prompt change, a model update or a data shift degrades quality *without a single test turning red* — unless you built the instrument that notices. This lesson covers building an eval suite from your real behaviour requirements, measuring retrieval and generation separately, the three budgets every AI feature has (money, latency, quality), instrumentation for all three, and the optimisations that move them — caching, smaller models by task, batching. You cannot improve what you do not measure, and you cannot ship what you cannot afford.

3. Safety, privacy and shipping

Shipping AI means deciding what it may never see, never say and never do — and making those decisions in code and policy rather than hoping the prompt holds. This lesson is the pre-launch pass: harm scenarios and their mitigations, data protection across the whole pipeline, content moderation and output filtering, human oversight for consequential actions, incident response for AI-specific failures, and the audit and policy artefacts stores and regulators now expect. Safety is a set of engineering decisions with owners, tests and dates — which is what makes shipping defensible.