ReviewIQ

Turn reviews into insights. AI extracts what matters.

Comment

Inspiration

Online shoppers and small merchants face thousands of product reviews. Reading them all is impossible, but skipping them means missing patterns—what people love about fit and fabric, and what keeps coming up as complaints like shrinking or wrong sizing.

Human user: A shopper deciding whether to buy, or a merchant improving a clothing line.

Problem: Reviews are useful but overwhelming; signal is buried in noise.

What it does

ReviewIQ lets you paste or upload clothing reviews and get an insights dashboard in seconds:

  • Sentiment breakdown — positive, negative, and neutral percentages
  • Top features loved — ranked phrases like "soft fabric" and "good fit"
  • Top complaints — ranked issues like "shrinks after wash"
  • Trending words — word cloud colored by sentiment
  • One-line summary — e.g., "72% positive; most loved: soft fabric; main complaint: shrinking"
  • Export — PDF or JSON for sharing with a team

The human stays in control: ReviewIQ summarizes; they decide what to buy or fix.

How we built it

Architecture

  • Frontend: Next.js 14, React, TypeScript, Tailwind CSS, Recharts (Vercel)
  • Backend: Node.js, Express, TypeScript (Render)
  • AI: Hugging Face Inference API — facebook/bart-large-mnli for zero-shot sentiment (positive / negative / neutral)
  • Insights: Domain regex patterns for clothing features and complaints + keyword frequency for trends

Data

  • Women's e-commerce clothing reviews (~23k dataset, Kaggle)
  • Preprocessing: scripts/process_reviews.py (clean text, stopwords, CSV column detection)
  • Live demo uses pasted/CSV samples; pipeline validated on dataset samples

Key decisions

  • Pre-trained zero-shot instead of fine-tuning — reproducible, fast to ship in 24 hours
  • Rule-based fallback when HF API unavailable
  • Aggregate insights only — no scoring individual reviewers

Links

Challenges we ran into

  • Sarcasm and slang — zero-shot can misread ironic praise; we document this in About and Model Report
  • Cold starts — Render free tier sleeps; we wake the API before demos
  • Rate limits — batch sentiment calls with small delays; cap reviews per request for reliability
  • Novel phrases — regex features miss new wording; trade interpretability for coverage

What we learned / What's next

  • Augmentation works best when the UI shows why (ranked features + complaints), not just a single score
  • Next: fine-tune on the full 23k dataset, category-level sentiment, and merchant alerts for rising complaints

Responsible AI

ReviewIQ helps shoppers and merchants see patterns in reviews; it does not replace human judgment. People harmed by mispredictions include merchants who over-trust a wrong summary and shoppers misled by biased or sarcastic reviews misclassified as positive. We are out of scope for automated product delisting, hiring/firing, or identifying individual reviewers. Guardrails: minimum 10 reviews per run, aggregate outputs only, documented limitations (sarcasm, slang, review bias), About-page ethics notes, and no PII in uploads. Users should verify insights against original reviews before business decisions. ReviewIQ is decision support—the human remains the driver.

Built With

Share this project:

Updates