Build reliable AI systems,
from prompt to production.
A practical engineering curriculum for building, evaluating, securing, and operating modern LLM applications.
28 modules · 27 exercises · Learn at your own pace
01 Design prompts with measurable contracts
02 Build secure retrieval and tool workflows
03 Evaluate quality, latency, and cost
04 Ship observable production agents
Follow concrete engineering failures from a soft prompt contract to an evaluated, authorized, observable service. Each core module opens with an incident, asks you to make the decision that closes it, and returns to the evidence that would prove the fix. Cases stand alone, so shorter learning paths and search arrivals still work.
Choose the route that fits your goal¶
Whether you have a weekend or want a complete engineering foundation, begin with a focused route and expand when you need to.
Build your first LLM app
Learn prompt contracts, basic retrieval, and evaluation by shipping a small working application.
Follow the Weekend path → Most complete 02 8–12 weeksBecome a production AI engineer
Build dependable services with security boundaries, eval gates, observability, and cost controls.
Follow the Professional path → 03 90-day projectsDevelop a specialization
Go deep on a stock recommender, hybrid neural models, or an agentic editor plugin.
Explore specialization tracks →A curriculum built around five engineering gates¶
The modules move from dependable model interactions to complete production systems. Each gate is a working-system exit criterion — the previous gate's failure mode is what forces the next one.
Make model behavior explicit, schema-valid, and safe against hostile input.
Trustworthy output is unmeasured output until it has an eval suite.
Give models the right context, and prove retrieval is even needed.
Design workflows that can act without losing control or budget.
Survive real traffic, audits, drift, and a bill someone has to explain.
See exit criteria for each gate →
View the complete capability map →
Prove all five gates on one system¶
Three capstones. Each starts from a runnable repository with planted holes, and each closes the same five gates on a deliberately different shape of system.
One request path, end to end: a model classifies a ticket, policy evidence gates a refund, and an authenticated human approves it without double-charging anyone.
A 4B model may propose a patch; deterministic checks decide whether that patch is legal. The specification is frozen before any code is generated.
A task a 20B model fails as one prompt, finished by splitting the work, solving each part, and joining the parts under checks you wrote first.
Learn by building, not by collecting vocabulary¶
Understand the system
Incident stories and visual mental models show where data, trust, and tokens move.
Make the tradeoff
Focused questions turn abstract concepts into concrete engineering decisions.
Prove it works
Labs, tests, and interactive quizzes end in a small artifact you can actually ship.
Start with a working environment.
Set up the repository, run your first examples, and begin Module 01.
Progress and XP stay in your browser via localStorage. No account, analytics service, or backend is required.