Search⌘ K
AI Features
Home
Courses
Forward Deployed Engineer for AI Systems

Capstone: LLM Application Challenge

Complete a timed, independent assignment to build a policy assistant that grounds answers in retrieved documents and uses a bounded review-draft tool only after explicit confirmation. Evaluate retrieval quality, answer correctness, and tool outcomes separately. Prevent instructions in retrieved content from overriding application rules, and distinguish fixture-based test results from verified live execution.

We'll cover the following...

This capstone is a second, independent timed assignment. This time, the customer needs an LLM-powered assistant:

“I search our policy notes and record unresolved questions for finance by hand. Give me a local assistant that answers expense-policy questions with sources I can check, and lets me confirm when an unresolved question should become a review draft.”

Expense policy coordinator

Starting the timed attempt

Confirm that you are ready to begin before opening the customer brief. Use a separate project package and begin without reusing the work-order capstone implementation.

Start-of-attempt notice: Opening the full brief and fixtures starts the 240-minute active-work limit for your first attempt. Reading, implementation, evaluation, documentation, and recording all count toward this limit. When the limit expires, freeze the submission, document unmet requirements, and accurately report the time spent.

For this independent course attempt:

  • You may use standard documentation and software libraries.

  • AI coding or writing assistance, implementation help from others, and annotated model solutions are prohibited.

  • You may use a language model as part of the application's runtime, but not to generate implementation code or submission materials.

  • Do not open the annotated reference until you have frozen your initial submission and recorded your self-score.

These restrictions apply to this course exercise. Follow the current instructions for any employer assessment. ...

Declare Your Readiness
Select all that apply

Q1. Before opening the full brief, select every readiness declaration that would be permitted under the rules. The three compliant declarations are alternatives, not requirements to start. This records your declaration only; it does not start or track a clock or verify credentials.

Do you find this helpful?

This declaration does not start or track your clock or verify credentials. Track active work yourself when you open the full brief and fixtures. If runtime access or budget is blocked, fixtures can test containment, but they cannot establish how a live model answers.

The customer brief

...