Independent class project · Agentic AI final

What are app reviewers complaining about, and what should be fixed first?

Loading data from the database…

AI recommendation

AI-generated

Click a C1 chip to see the number behind it, or a review ID to read the review.

Alternatives considered

    Memo's stated limitations

      Issue ranking

      priority score = sum of severity

      Complaints and cancellation threats grouped into issues. Select a row for its trend and evidence.

      #Issue ComplaintsMean severity Priority score

      Complaint share by month

      Complaints + cancellation threats ÷ all classified reviews that month. Hover for the counts.

      Show as table

      Topics

      All classified reviews by primary topic.

      Severity

      1 = no problem … 5 = serious financial, privacy or data harm.

      Intent

      How this was made & limitations

      1. Ingest: every public review in the 18-month dataset is accounted for; empty or failed rows are quarantined with a reason, not dropped.
      2. Enrich: a language model labels each review's topic, intent, severity, sentiment and an exact evidence quote (up to 50 reviews per request; identical texts reuse a saved result).
      3. Verify: a second model re-checks a sample, and labels are compared with 50 hand-labelled “golden” reviews.
      4. Group & rank: complaints are grouped into issues and ranked by total severity using plain code, not the model.
      5. Recommend: a model writes the memo from the saved ranking; every number it cites is a claim in the database.
      • Reviews are self-selected and public; they over-represent strong opinions and are not a sample of all users.
      • An expressed intent to cancel is not observed churn.
      • There is no revenue, usage or account data, so the ranking measures reported pain, not business impact.
      • Labels come from a model and can be wrong; see the golden-set agreement above.

      Review

      Loading…