GUIDED PROJECT BEGINNER PRODUCT ANALYTICS

App Review Triage what should we fix first?

Roadmap lock is Thursday and there are 3 engineers, not 30. Everyone's got a theory: ads, buffering, Shorts. You'll classify 2,170 real App Store reviews with the Claude API, rank the complaints, filter to what's actually fixable, and ship a 4-slide recommendation deck.

TOOLS & TECHNIQUES

Python pandas Matplotlib Claude API Prompt engineering

▶ Start chapter 1

Google Meta

I've personally interviewed hundreds of data candidates.

Pasquale Bruno Founder · datascienceportfol.io Read bio

⏱

5–6 hours

self-paced

◉

6 chapters

video + written

◱

Real dataset

5,695 App Store reviews, 12 storefronts

✦

Built with Claude

Ch 3: 2,170 reviews classified; Ch 4: recommendation drafted

→

Portfolio deliverable

Repo + 4-slide deck + LinkedIn post

WHAT YOU'LL WALK AWAY WITH

Real, portfolio-ready artifacts, not a certificate.

✓

Analysis notebook

✓

Enriched, classified dataset

✓

4-slide stakeholder deck

✓

GitHub repo with README

✓

Portfolio case study page

THE SCENARIO

The biggest complaint and the biggest job are different things.

**Priya** is a Group Product Manager with three engineers for the quarter and a roadmap meeting on Thursday. Everyone on her team has a theory about what's broken (ads, buffering, Shorts), including her own. She needs to know what people are actually complaining about, and whether it's the same everywhere.

You'll classify 2,170 real YouTube iOS App Store reviews with the Claude API, build a comparable window across five markets after catching a data-collection trap that would have wrecked the comparison, and filter the complaints down to what three engineers can actually fix. The biggest complaint, ads at 33%, turns out not to be the answer. Crashes are 78% of everything blocking and actionable.

BY THE END, YOU'LL BE ABLE TO

✓

Catch a confounding trap before it reaches a slide, and know the general habit that prevents it

✓

Write a constrained-output Claude prompt that turns free text into countable categories

✓

Validate a model's classification output rather than trusting it because it ran without an error

✓

Choose the right denominator for a metric, and change it deliberately rather than by accident

✓

Turn a ranked list of complaints into a recommendation scoped to a real constraint: three engineers

PROJECT CURRICULUM

6 chapters · ~5–6 hours

CLAUDE API NEW 16 videos 6 written lessons

Project brief & setup

video + reading

40 min +

● Read Priya's brief: 3 engineers, a Thursday deadline, and three competing theories about what's broken

● Set up Colab, Google Drive, and a GitHub repo with the right folder structure

● Store the Claude API key safely and confirm it works on a real review

● Write a five-question pre-analysis plan before looking at any data

Meet the data

video + code

40 min +

● Profile 5,695 reviews: no nulls, a bimodal rating distribution, 31% under 30 characters

● Find that the mean rating (3.26) describes almost nobody in a distribution where 77% sit at the extremes

● Notice the file spans March 2023 to July 2026, despite Apple's feed only serving 'recent' reviews

● Save a profiled working file, setting up the trap the next chapter falls into

Is it the same everywhere?

video + code

45 min +

● Run the obvious two-line comparison: mean rating by country (a 1.07-star spread)

● Discover it's confounded: the US's 500 reviews span one day, Kenya's span 1,226 days

● Build a comparable 3-week window across 5 markets (2,170 reviews) and re-run it

● Corrected spread drops to 0.47 stars: markets aren't meaningfully different, once compared fairly

Claude API: extraction 🧠 CLAUDE API

video + code

1h 0m +

● Write a constrained-output prompt that turns each review into 5 structured fields

● Classify a stratified sample of 400 reviews with the Claude API, one call each

● Validate the output: completion rate, parse-error rate, distribution sanity, and reading rows by hand

● Compare your 400 against the full 2,170 to confirm the sample was big enough for the headline categories

What's broken: synthesis & the deck 🧠 CLAUDE API

video + code

55 min +

● Rank complaints by type (ads 33.2%, crashes 27.2%) and check all three theories in the room

● Filter to blocking + actionable and watch the ranking flip: crashes are 78.3% of what's fixable

● Track the crash trend by app version, not by week: falling, but ticking back up on the newest build

● Find that markets share the same ratings but different complaint mixes (India 38.9% crash vs US 19.1%)

● Use a second Claude call to draft the recommendation from your verified numbers, then build the 4-slide deck

Ship it & frame it for employers

video + reading

1h 0m +

● Write a README that leads with the counter-intuitive finding, not the obvious one

● Get Claude to review your README as a skeptical hiring manager, then fix what it finds

● Publish a LinkedIn post and a portfolio page without overstating what the project is

● Rehearse seven real interview questions about this project out loud

▶ Start chapter 1

PREVIOUSLY

Google Meta

YOUR INSTRUCTOR

Pasquale Bruno Founder · datascienceportfol.io

I spent 10+ years in data and product roles at Google and Meta, including hundreds of hiring-panel interviews, so I know exactly what makes a hiring manager keep reading. I design and review every guided project myself, using briefs and datasets modelled on real analyst work, not toy CSVs. If you get stuck, you're emailing me directly, not a support queue.

10+ yrs

IN INDUSTRY

500+

CANDIDATES INTERVIEWED

300+

PORTFOLIOS REVIEWED

See all projects Connect on LinkedIn

SKILLS YOU'LL PRACTICE

Tools & techniques

Python pandas Matplotlib Claude API Prompt engineering Confounding Stratified sampling Model output validation Denominator selection Stakeholder communication

PREREQUISITES

Before you start

Comfortable with Python (functions, dicts, loops) REQUIRED

Some pandas (DataFrames, groupby) REQUIRED

Free Google Colab, GitHub, and Anthropic API accounts (~$0.25 of API credit needed) REQUIRED

No prior AI/LLM experience needed. Chapter 3 covers prompt writing from scratch HELPFUL

Familiarity with basic stats (mean, distribution) HELPFUL

🧠 NEW · USE AI LIKE A REAL JOB

Ship analysis with the Claude API in the loop.

Real data teams aren't writing SQL one query at a time anymore. A whole chapter teaches you to wire Claude into a real analysis, the way analysts at Stripe, Anthropic and Notion actually work.

Two-call architecture

One call per review for structured extraction (Ch 3), one call over your aggregates for the recommendation (Ch 4), never mixed

The model reads your conclusions, not your data

The synthesis call only sees numbers you've already verified, so it can't invent a finding

Built to survive being ignored

Claude wraps JSON in markdown fences despite being told not to. The prompt handles that, because an instruction is a strong request, not a guarantee

Pricing

One plan gets you every Guided Project.

Free builds your portfolio. Job Ready adds every Guided Project on top, live now and every one we ship next.

Free

$0/forever

For anyone with projects ready to share.

✓ Portfolio website with unlimited projects
✓ Personalised URL
✓ No design nor technical skills needed
✓ All your education and experience info in one place

Start free

Job Ready

$99/ month

For people investing in a data career.

✓ Portfolio website builder PRO
✓ All current + future Guided Projects
✓ Discord support channel for any questions, answered by me
✓ Weekly live office hours with me
✓ 1:1 monthly call with me (first 20 members)
✓ Async portfolio & LinkedIn reviews (first 20 members)
✓ Cancel anytime

Get Job Ready

Not sure if it's right for you?

Book a free 15-minute call with me. We'll talk through where you are and whether a Guided Project makes sense.

Book a call →

HOW WE'RE DIFFERENT

Why this isn't another course.

FEATURE THIS PROJECT DATACAMP COURSERA

Real business scenario, not toy data ✓ ✕ ✕

Portfolio-ready deliverable ✓ ✕ –

Real Claude API extraction pipeline, not a toy prompt ✓ ✕ ✕

You validate model output, not just trust it ✓ ✕ ✕

Certificate of completion ✓ ✓ ✓

Thousands of micro-lessons ✕ ✓ ✓

FREQUENTLY ASKED

Questions, answered.

Do I need AI/LLM experience? +

No, Chapter 3 teaches prompt writing from scratch, including how to handle a model that doesn't follow instructions.

Is the dataset real? +

Yes, 5,695 real public App Store reviews for YouTube's iOS app, across 12 storefronts, with the collection script included in the repo.

What does the Claude API cost? +

About $0.22 total for the whole project, roughly 400 extraction calls plus one synthesis call, on Claude Haiku 4.5.

Can I cancel anytime? What about refunds? +

Yes, cancel anytime, no lock-in, right from your account. We don't offer refunds on subscription payments already made, but you'll keep full access through the end of the billing period you've already paid for.

Can I use this in interviews? +

Yes, the project ships with seven interview questions and model answers, including the one about the confounding trap in Chapter 2.

CONTINUE YOUR PATH