NutriProof

Validating a claim about the cholesterol in eggs
Successful fine tuning of the model

Inspiration

NutriProof was born from our frustration with the overwhelming amount of contradictory nutrition information online. After witnessing friends and family members fall victim to dangerous nutrition misinformation, we realized how difficult it is for the average person to separate fact from fiction in health claims. We wanted to create a tool that could harness the power of AI and computational verification to give people instant clarity about the accuracy of nutrition information they encounter online.

What it does

NutriProof is a Chrome extension that acts as your personal nutrition fact-checker. Users can select any text containing health or nutrition claims, right-click, and choose "Check with NutriProof." Our fine-tuned AI system identifies factual claims, converts them into queries optimized for computational verification, and processes them through Wolfram Alpha to check their accuracy. Results are presented with an intuitive grading system (A-F) and interactive pie chart visualization, showing users exactly which claims are true, false, or somewhere in between.

How we built it

We built NutriProof using a combination of front-end and back-end technologies. The Chrome extension uses HTML, CSS, and vanilla JavaScript with Chart.js for visualizations. The backend runs on a Python Flask server that handles the AI processing and API calls. We fine-tuned GPT-4o specifically for nutrition claim identification and verification, creating a model that understands the nuances of health information. The system connects to the Wolfram Alpha API for computational verification against scientific data, with specialized query optimization for nutrition facts.

Challenges we ran into

Fine-tuning the AI to accurately identify implicit health claims was our biggest challenge. Many nutrition claims are presented casually ("Celery juice detoxifies your liver") rather than as explicit factual statements. We also struggled with optimizing Wolfram Alpha queries to handle nutrition-specific information, as many claims required reformulation to be computationally verifiable. Additionally, balancing accuracy with speed was difficult—users expect instant results, but the dual-API architecture initially created latency issues we had to optimize.

Accomplishments that we're proud of

We're most proud of our specialized fine-tuning process that enables the AI to understand nutrition claims with remarkable accuracy. The grading system we developed goes beyond simple true/false classifications to reflect the nuances of nutrition science. We also created a clean, intuitive UI that makes complex verification results easily understandable to non-experts. Finally, we're proud of the animation and visualization techniques that keep users engaged while providing meaningful information at a glance.

What we learned

This project taught us how combining specialized APIs with fine-tuned AI models creates capabilities greater than the sum of their parts. We learned about the challenges of nutrition science, where claims often exist in shades of gray rather than black and white. The development process deepened our understanding of Chrome extension architecture and the importance of user experience in technical tools. Perhaps most surprisingly, we discovered how many "common knowledge" nutrition facts completely fall apart under computational scrutiny.

What's next for NutriProof

We're planning to expand NutriProof's capabilities to analyze images of nutrition labels and ingredient lists. We're also working on a feature to provide contextual explanations for why certain claims are verified or debunked, helping users build their nutrition literacy over time. In the longer term, we aim to create a community-driven database of verified nutrition claims to speed up the verification process and reduce API calls. Finally, we plan to release a mobile app version to make nutrition fact-checking accessible across all platforms.