Build the four no-code artifacts and two scoring sheets that get you in the room — then circle back for the full 12 weeks
The full track is sequenced by topic, so the strongest portfolio pieces are spread across the weeks. This lane pulls them to the front. Do these six steps — mostly writing, a little spreadsheet work — and in about two weeks you have a portfolio you can show. Then run the full track for the depth that makes you defend it in an interview.
Weighted dimensions (prompt adherence, aesthetics, technical/artifacts, text rendering), a safety gate, anchored scale points, worked examples + counter-examples, and an edge-case section. This is the single highest-signal artifact you can own.
A clear, mutually-exclusive set of failure categories for image/video outputs, each with a severity tier and 1–2 examples. This is what turns "it looked bad" into structured, trackable data — and it's the backbone of every QA report.
A checklist reviewers run on each item, plus the QA-sampling routine you'd use to spot-check a batch (sample rate, gold items, what gets escalated). Demonstrates you can keep a team's output consistent.
How you bring a group of reviewers to agreement: calibration-session cadence and agenda, what you review, how disagreements get resolved and fed back into the guidelines. This is the "lead" in Human Data Lead.
A spreadsheet that scores a batch of generations against your rubric: columns = dimensions, rows = items, formulas = weighted score + verdict + weakest dimension. No code required — Sheets formulas are enough.
A one-page report comparing two models on a held-out prompt set: win rate, per-dimension breakdown, example wins/losses, and a recommendation. Pair it with a note on contamination and position bias so it reads as rigorous.
Do Kaggle Learn Python + Learn Pandas (~a day), then recreate your Scoring Sheet from a CSV in Pandas and reproduce a reviewer-agreement number. That single move lets you say you do eval QA in both Sheets and Python. Week 7 walks it →