b7e21d9branch(build):git switch app-engineer
App Engineer
Ship an AI feature with types, tests and a fallback, not a demo that only works on stage.
Treat the model like any other unreliable dependency. You wrap it in contracts, give it real data through retrieval, and keep a test set that tells you when a change made things worse.
- for
- Web and app developers who want AI features that behave like the rest of their code.
- time
- 4 weeks at 3 to 4 hours a week
- not for
- People who want a no-code tool. Launchpad is the better start.
You will ship
A help-centre answer box that cites the exact paragraph it used and says so when it does not know.
Prerequisites
- Comfortable in one language (TypeScript or Python in the examples)
- Has called a REST endpoint before
- Basic git
b7e2e01plan:six commits to ship
The branch, commit by commit.
0/6 commits done · saved in this browser only
- commit 1Typed model wrapper
One function, typed input and output, timeouts and a logged request id.
Where this happens: /courses/llm-app-basics#m2 - commit 2Chunk and rank
Split documents, score them against a question, and look at what came back.
Where this happens: /lessons/rag-retrieval#try - commit 3Cite or refuse
Answers must point to a retrieved chunk. No chunk above the threshold means no answer.
Where this happens: /lessons/rag-retrieval#build - commit 4Twenty-question test set
Write questions with known source paragraphs and measure how often retrieval finds them.
Where this happens: /lessons/rag-retrieval#test - commit 5Fallback path
When the model is slow or down, the feature degrades to search results, not an error page.
Where this happens: /courses/ship-and-see-it#m3 - commit 6Ship the demo
Put it behind a flag for a small group and read every answer it gives in week one.
Where this happens: /projects#docs-answer-box
b7e2c02docs:courses on this branch
Courses.
LLM Apps, Properly
Call a model like a careful engineer: typed inputs, structured outputs, timeouts, streaming and logs.
Retrieval That Cites Its Sources
AI that searches your own documents first, then answers only from what it found, with the paragraph to prove it.
Evals Before Vibes
Build the test set first. Then every prompt change, model swap or vendor pitch gets judged on the same evidence.
b7e2d03ship:capstone
Capstone: Cited answer box
Answers help-centre questions from retrieved paragraphs, with a link to each one.
Tests it must pass
- refuses questions outside the docs
- recall at 3 >= 5 of 6
- every answer has a citation
if (!ranked.length || ranked[0].score < threshold) { return { answer: "I don't have enough material to answer that.", citations: [] };}