The Fantastic For: Unstoppable

Online reviews heavily influence how we choose restaurants, salons, and local businesses, yet they’re often polluted with irrelevant, misleading, or low-quality content. We were inspired to tackle this “trust in reviews” dilemma—how can platforms ensure that what users read is both relevant and reliable? Our goal was to design an ML-powered system that automatically evaluates the quality and relevance of Google location reviews, helping users make better decisions and allowing businesses to be judged fairly.

We built the system by combining feature engineering and modern NLP techniques. Using Python, Pandas, and Scikit-learn, we extracted handcrafted features such as sentiment ratios, text statistics, and restaurant-specific mentions, and combined them with TF-IDF embeddings. On the backend, we integrated Hugging Face models for zero-shot classification (semantic relevance) and sentiment analysis, alongside lightweight vision models for NSFW and promotional image detection. A Flask-powered backend feeds results into a clean, interactive web UI that shows users exactly why a review was approved, flagged, or removed.

Along the way, we learned the importance of balancing accuracy with transparency and usability. We faced challenges like noisy data, TF-IDF “empty vocabulary” errors, and latency trade-offs between heavy NLP models and real-time responsiveness. Solving these problems taught us that building trustworthy AI systems isn’t just about algorithms—it’s about explainability, user trust, and scalable design.

Built With

• css-frameworks-&-libraries:-flask-(backend-&-api)-pandas
• css3-(interactive-moderation-dashboard)-data-handling:-json
• csv
• devpost-(submission)-apis-&-models:-hugging-face-model-hub-(distilbert-base-uncased-mnli
• distilbert-base-uncased-finetuned-sst-2-english
• excel-(via-openpyxl)-cloud/platforms:-google-colab-(training-&-prototyping)
• for
• github-(version-control)
• html
• html5
• javascript
• languages:-python
• nsfw-image-detection-models)-tools:-jupyter/colab-notebooks
• numpy
• pip/virtualenv
• scikit-learn-(feature-engineering-&-ml-pipeline)-hugging-face-transformers-(zero-shot-classification
• sentiment-analysis)-nltk-(text-preprocessing)-opencv-/-lightweight-vision-models-(nsfw-&-promo-image-detection)-frontend/ui:-vanilla-js