Roundabout

Inspiration

We wanted to solve the problem of misleading Google reviews that distort user decisions and harm businesses. Ads disguised as reviews, irrelevant comments, and fake rants make it hard to trust ratings.

What it does

Roundabout detects and flags reviews that violate three key policies:

Advertisements
Irrelevant Content
Rants Without Visiting
It cleans and processes data elegantly, ensuring higher-quality insights for businesses and users.

How we built it

Generated pseudo-labels using GPT-4o and fallback with Ollama Gemma3-4b
Trained and fine-tuned a BERT-based classifier because of its strong contextual understanding of language, making it ideal for nuanced review classification.
Built a UI for seamless interaction and hosting the model.

Challenges we ran into

Token limitations when using GPT for pseudo-labeling
Prompt engineering mistakes that led to incorrect labels
Classifiers producing inconsistent & inaccurate results

Accomplishments that we're proud of

Building a smooth, end-to-end AI pipeline in just 24 hours
Successfully integrating LLM-driven pseudo-labeling with a custom-trained classifier
Hosting the model on a functional UI for real-time predictions

What we learned

The importance of data quality in training AI models
How prompt engineering impacts the reliability of pseudo-labels
Practical deployment challenges for machine learning systems

What's next for Roundabout

Improving the trained classifier to increase accuracy and reduce false positives
Expanding categories and policies for broader review moderation
Exploring multilingual support for global use cases

Due to the lack of time & resources, we were unable to effectively train our classifier to be less strict on policy detection. This led to a high rule out rate & false positives for the advertisements. However, our results are promising and we hope to continue developing on it to improve.