← All projects
Project

App Review Pulse

An AI agent that turns hundreds of app-store reviews into a one-page weekly note, delivered through MCP.

Role
Solo, from a brief
When
Jul 2026
Status
Working pipeline
Built with
Python, Gemini, Groq, MCP
The obvious version

A sentiment dashboard

What I built

A 250-word weekly note in Docs and Gmail

Left out

More than five themes Reviewer names

It's working if

1+ backlog action per note

01Where it started

App-store reviews are public, plentiful and noisy. Product, support and leadership teams either ignore them or lose hours reading them by hand. The usual answer is a dashboard of sentiment scores.

02What changed my mind

A dashboard waits to be opened, and in a busy week nobody opens it. What the team needs is the reading already done: what people are unhappy about, in their own words, and what to do about it, delivered where they already work.

How strong is this? A written note came from the brief. Capping it hard, at three themes, three quotes and three actions, was my call.

03What I built

A pipeline that imports 8 to 12 weeks of reviews, groups them into at most five themes, and writes a note of 250 words or fewer with the top three themes, three verbatim quotes and three action ideas. A second model tightens the writing, then the note is published to Google Docs and drafted as an email in Gmail.

Docs and Gmail are reached through MCP servers, so the agent never handles Google credentials directly. Names, emails and device IDs are stripped before anything is written.

04What I left out, and why

  • A dashboardPeople only pull information when they have time. A note gets pushed to them.
  • More than five themesA long list of themes is a dashboard in disguise. Five is the cap, and the note shows three.
  • Reviewer namesQuotes are anonymous. The note gets shared widely, and no reviewer agreed to that.

05How I'll know it's working

Main metric target

1+

action from each weekly note makes it onto the backlog. A note nobody acts on is just a nicer way to ignore reviews.

What moves it
  • Share of recipients who open the note
  • Themes that repeat week after week, a sign they aren't being fixed
What must not get worse

No personal data in any note.

06What I got wrong

My test run used only 10 reviews, which is too few to trust the themes. The next run needs a few hundred real reviews before the grouping means anything.