Orientation · Read First

How to Use This Track

The bar you're aiming for, the order that works, and the portfolio that gets you hired

How much Python do you actually need?

Less than this track might make it look. The 12 weeks go deep so you understand the why — but for a Human Data Lead role, Python is a credibility booster, not a requirement. You are not here to build models. You're here to lead human reviewers and be fluent enough in the data to keep them honest.

Your real target — say this in an interview

“I'm comfortable using Python and Pandas to inspect evaluation results, analyze reviewer agreement, identify failure patterns, and support QA workflows.”

That's the level. Not “I build models.” If you can do that from a CSV, you've cleared the bar.

The Whole Toolkit

Learn (this is enough)

You do not need

Why the track still goes deeper

The weeks on RLHF, reward models, and DPO aren't there to make you an engineer — they're so you can talk to researchers as a peer and design the data they need. Understand the concepts; you do not need to implement any of them. SQL (Week 7, Day 34) is a nice-to-have for credibility, not a requirement — skip it if time is tight.

The order that works: artifact → Sheets → Python

Don't start with code. Start with thinking. The most valuable artifacts in this field are documents and decisions, not scripts. Build in this order and you're never blocked on syntax:

  1. Build the artifact with no code. A rubric, a failure taxonomy, a calibration guide, a QA checklist — these are written documents. They demonstrate the core skill and need zero Python.
  2. Do the scoring in Google Sheets. A prompt-adherence scoring sheet or a model-comparison report works perfectly as a spreadsheet first. Columns are fields, rows are items, formulas are your stats. This is how a lot of real eval QA actually happens.
  3. Then recreate it in Python/Pandas. Load the same data from a CSV, reproduce the analysis, and now you can say you do it both ways. That's the credibility layer — added last, on top of work that already stands on its own.
Practical learning path for the code part

Do Kaggle's Learn Python and Learn Pandas (free, in-browser, ~a day total) — then stop. Resist going further. The goal is to recreate a scoring analysis from a CSV, not to become a software engineer.

Your portfolio — the artifacts that get you hired

These are the deliverables to walk into an interview with. Build the no-code ones first; they're the strongest signal. Each maps to the week where this track teaches it.

ArtifactBuilt inTool level
AI Video Evaluation RubricWeek 2 + Week 5No-code
Failure Taxonomy (categories + severity for video/image failures)Week 3 (+ Week 6, 9)No-code
Human Feedback Calibration GuideWeek 10No-code
Reviewer QA ChecklistWeek 9–10No-code
Prompt Adherence Scoring SheetWeek 5Sheets → Python
Model Comparison ReportWeek 6Sheets → Python
Reviewer Agreement Analysis (kappa, disagreement hot-spots)Week 3 + Week 7Python (Pandas)
Quality / Drift AuditWeek 9Python (Pandas)
Build these four first — they need no code at all

AI Video Evaluation Rubric · Failure Taxonomy · Human Feedback Calibration Guide · Reviewer QA Checklist. Done well, these alone make a strong portfolio. Then add the two Sheets pieces, then the Pandas layer on top.

The hiring story you're assembling

A hiring manager seeing this combination reads it as “can lead human reviewers and understands the data side” — which is exactly the Human Data Lead profile:

The combination

Emmy-winning video / post-production lead  +  xAI / Micro1 evaluation experience  +  rubric design  +  reviewer calibration  +  basic Python / Pandas for evaluation QA.

Your production background is leadership credibility most data candidates don't have. Your tutoring is hands-on evaluation experience. This track adds the two things that complete the picture: the methodology (rubrics, agreement, calibration, QA) and just enough data fluency to be dangerous. Everything here serves that story — don't let the depth distract you from it.

The One-Paragraph Plan