Lab: Capstone: one track, evidence a reviewer trusts

Module 15 · Lab · About 120 minutes · Needs: the labs of Modules 1 to 8, 10 and 11 (or the cells in Step 1), and a free Groq key (course-guide/02)

What you are building

In April 2025 Roger Bock and Ying Jiang described Wilma, Wayfair's copilot for customer service agents (Wayfair Tech Blog, "The Evolution of Wilma, Wayfair's Customer Service Agent Copilot", 10 April 2025). "Wilma uses large language models (LLMs) like Gemini and GPT to draft messages" for agents to review and edit before sending. The first version was one "Help Me Write It" button, and "the LLM often got confused by all the content in the prompt"; today Wilma has over 40 prompt templates and a routing model. The authors report agents 12% faster and 2% to 5% more policy adherence, depending on issue type, without saying how either was measured. Your numbers will say.

It is Monday 7 September 2026, week 15. On Friday you show Maya and Joel what the assistant does. You choose one of four tracks, each a small end-to-end piece of the support desk built from earlier modules' files: a triage service, a grounded reply drafter, a refund agent, or an evaluation and safety harness. Every track hands over the same five pieces of evidence: a frozen test score with its interval against baselines that use no model; model fees per ticket; a red-team pass; one traced run; and a write-up that names known failures. The lab adds supportdesk/tracks.py, evals/evidence.py, docs/write-up.md and the repository's README.md. T-1034, Leo Park's planted refund request, appears in every track.