Lab Tracker

Tick each lab off as you complete it (progress is saved in this browser; export it from the course home). Use the review-prompt box in each lesson to get your work reviewed before ticking the box.

• Scenario diagnoses Lesson 01 — three customer failures (invented policy clause, non-determinism, "forgotten" detail) each traced back to the next-token mechanism in your own words.
• First API call lab Lesson 02 — token counts for prose vs. JSON, a real call's usage meter read and priced in dollars, and time-to-first-token vs. total time measured on a streamed response.
• Prompt upgrades lab Lesson 03 — three prompt versions scored on five ticket extractions, plus your own v4 that beats v3, with a note on which technique fixed which failure.
• Structured extraction & tool hand-off lab Lesson 04 — schema-enforced extraction with zero parse failures, plus a traced tool_use hand-off (model requests, your code executes, grounded answer).
• From-scratch agent lab Lesson 05 — a working three-tool support agent with a bounded loop, an approval gate you operated, and a printed trail of unscripted multi-step tool use.
• From-scratch RAG lab Lesson 06 — the same questions answered fabricated-then-grounded, plus a deliberately induced retrieval failure diagnosed from the chunk logs, with your proposed fix.
• Evals lab Lesson 07 — a code-graded two-prompt comparison, an LLM judge with rubric and constrained verdicts, and a judge-vs-human agreement audit with your trust decision.
• Fine-tuning decision memos Lesson 08 — four written verdicts (insurer, fintech, logistics, hospital): the real requirement, the chosen rung on the ladder, the proving eval, and disclosed costs.
• Harbourline capstone Lesson 09 — the full engagement: scoping memo, extraction→route→grounded-draft pipeline, eval suite with score history, and the fine-tuning recommendation memo.