Course
LLM Application Engineering
Build software on top of language models like an engineer, not a demo: prompting as an interface, structured output, RAG that retrieves the right thing, tools and agents with boundaries, evals, and the cost math that decides whether it ships.
Foundations
What a model actually is to your application: a stateless text function with a context budget and a temperature dial.
- 01Prompting Is API DesignTreat the prompt as a contract: specify inputs, outputs, and failure modes like you would for any boundary.20mOpen
- 02Tokens and Context: The Real ConstraintsContext is a budget, tokens are the currency — and the model reads the middle of a long context worst.18mOpen
- 03Structured Output: Making Models Return DataJSON mode, schema enforcement, and the parse-retry pattern — turning prose machines into functions that return types.20mMembers
Retrieval & Tools
RAG that retrieves the right chunk, tool calling with real error handling, and where agent autonomy should stop.
- 01RAG That Retrieves the Right ThingRetrieval is the product — chunking, embeddings, hybrid search, and the eval that tells you retrieval is actually working.25mMembers
- 02Tool Calling and Agents: Autonomy With GuardrailsFunction calling done right: narrow tools, validated args, bounded loops, and a kill switch on anything with side effects.25mMembers
Production LLM
Evals before vibes, guardrails that hold, and the latency/cost math that decides whether a feature ships.
- 01Evals: Testing the UntestableYou can't unit-test prose, but you can regression-test behavior — fixture sets, model-graded evals, and CI that catches prompt regressions.22mMembers
- 02Cost, Latency, and GuardrailsThe economics and safety layers of a shipped LLM feature — caching, routing, budgets, and what to do when the model is wrong in forums.20mMembers