1. Courses
  2. AI Systems for Builders
  3. Capstone

Capstone project for AI Systems for Builders

Design and prototype a store-staff assistant

Mangrove Market wants an assistant in its Store app that answers staff questions from about 40 SOPs, the step-by-step guides for jobs like receiving deliveries, counting stock and handling customer returns. Store managers and cashiers may read different SOPs, and the assistant may look up the status of a goods received note (GRN) for the user's own store. About 320 stores ask roughly 2,000 questions a day, four times that during festival rushes, and many staff work on patchy store Wi-Fi. Write a short design, build a small working prototype of retrieval and the API client in JavaScript, and set up the repository rules so that AI-written code for this feature is reviewed and tested properly.

What you hand in

  1. A design document of 1 to 2 pages: the ingest and query paths, where permissions are enforced, the tools the model may call, the fallback when the model API is down, and a monthly cost estimate with your assumptions
  2. A retrieval prototype in JavaScript: 10 to 20 SOP chunks with small made-up vectors, cosine similarity, top-k with a role filter and a minimum similarity, and an eval set of at least 12 cases with its pass rate
  3. A client wrapper (it can call a fake API) that streams text, retries only retryable errors with backoff and jitter, honours retry-after, never restarts after text is shown, and logs request IDs and cache usage
  4. An AGENTS.md and a review checklist for the assistant's repository, plus a CI rule that blocks source changes without tests

Check your work

Tick each item when your project meets it. Be strict: this is the rubric the reviewer uses.

  • Permissions are enforced in the retrieval query, and a cashier can never receive a manager-only chunk, which an eval case proves
  • Weak matches return an honest "not found" instead of calling the model
  • The eval set includes out-of-scope, permission and prompt-injection cases, not only happy paths
  • Every tool runs with the user's own permissions, and any tool that changes data returns a proposal the user confirms
  • The prompt puts stable content first and per-request details (date, store, user) after the cache breakpoint
  • The client has one retry layer, a total time budget and a first-token timeout, and stops on spend-cap and 4xx errors
  • The cost estimate shows its formula and assumptions, including festival-rush volume and cache hit rate
  • The repository rules name the exact test command, forbid new dependencies without a reason, and require review for payments, auth and CI files

0 of 8 checked

Model answer

The model answer opens once your capstone passes the review, so you can compare your work with an outline of a strong submission.

Hand in your capstone

Your links are checked automatically, then the work is scored against the rubric above, item by item, with comments. A pass counts toward the course certificate and shows on your profile.

Loading…