Week 7 of 12 · The Lead

From Numbers to Recommendation

Lock in the data stack, then learn the move that separates a lead from an analyst: turning a finding into a decision someone can act on

Day 35 ~60 minutes Review

Day 35 of 66

What you now know

One week ago, "is our data good?" meant pulling someone aside. Now you can answer it yourself, with a table. Say the week's spine back without looking:

The Week 7 spine

The move that makes you a lead

Here's the trap. You now can make a beautiful table. An analyst stops there — "here are the numbers." A lead goes one step further: every finding ends in a recommendation. Not "annotator C's mean dropped 0.6," but "annotator C's mean dropped 0.6 on safety items only — recommend a 20-minute calibration on the safety guideline before next batch."

Finding → recommendation, every time

A chart is a question; a recommendation is an answer. The pattern is always: finding (what the data shows) → so-what (why it matters) → recommendation (the specific action) → how we'll know it worked (the metric to recheck). If your analysis doesn't end in an action someone can take this week, you've done the analyst's half of the job, not the lead's.

Self-quiz (no notes)

Answer out loud or in writing. If any answer is shaky, reread that day before moving to Week 8.

  1. You have labels.csv. In one line of pandas, get each annotator's count, mean, and std.
  2. The interesting breakdown is "score by task type," but task type is in a different table. What operation do you need, and on what?
  3. Annotator C averages 2.7 over 9 labels. Why shouldn't you act on that yet?
  4. The data is 8 million rows in a warehouse. pandas or SQL — and write the GROUP BY skeleton.
  5. Your audit shows quality dropped the same week you changed the routing rules. What's the one question to ask before blaming the change?
Answer key (peek after you try)

1) df.groupby("annotator")["score"].agg(["count","mean","std"]). 2) A join/merge on item_id to pull task type onto the label rows, then group by it. 3) Sample size — 9 labels is within sampling noise; you need the count beside the mean before trusting it. 4) SQL: SELECT annotator, COUNT(*), AVG(score) FROM labels GROUP BY annotator. 5) "What else changed that week?" — correlation isn't causation; rule out confounders (new annotators, harder items) before crediting the routing change.

Close the week

Wrap-up

Before Week 8

The reusable notebook is the real artifact

Don't throw away the analysis from this week — package it. A single notebook (or script) that takes any weekly label export and spits out the profile, the hot-list, and a chart is the tool you'll run every Monday as a lead. Build it once now; you'll reuse it for the rest of the track and into the job. Feed it next week's pipeline output and it keeps paying off.

The lead move

An analyst hands you the table and stops: "here are the numbers." A lead carries every finding the last mile — finding → so-what → recommendation → how we'll know it worked — so the analysis ends in an action someone can take this week, not a chart that raises a question.

Say this in an interview: "I never end on a number. 'Annotator C's mean fell 0.6 on safety items, so I recommend a 20-minute calibration before the next batch, and I'll recheck agreement on those items to confirm it took.' The recommendation is the deliverable; the table is just how I found it."

Week 7 Takeaways