HalluciCheck

Inspiration

AI assistants are becoming increasingly useful for research, education, and everyday tasks, but they can sometimes generate information that sounds convincing while being inaccurate. We wanted to build a tool that makes it easier for users to identify these hallucinations and understand why a claim may be unreliable.

This led us to build HalluciCheck, an AI-generated content verification platform that extracts factual claims from AI responses and checks them against available evidence.

What We Built

HalluciCheck allows users to paste AI-generated content and run an analysis. The system identifies individual claims, verifies them using web-based evidence, and presents the results in an easy-to-understand dashboard.

The application provides:

Claim extraction and verification
Verified, hallucinated, uncertain, and disputed verdicts
Confidence scores for individual claims
Evidence and source information
Overall hallucination/risk scoring
Search and filtering of extracted claims
Detailed claim views with explanations and sources
A responsive and polished dashboard UI

How We Built It

The project uses a Next.js/React frontend for the user interface and a Python FastAPI backend for analysis and verification.

The frontend contains the submission interface, analysis results dashboard, claim tables, risk visualization, and detailed evidence views. The backend processes submitted text and performs claim verification using web search and evidence.

We connected the frontend and backend through an API and used structured JSON responses so that verification results could be displayed dynamically in the dashboard.

We also focused on creating a clean user experience with responsive layouts, reusable components, filtering, sorting, confidence indicators, risk visualization, and clear verdict labels.

What We Learned

Through this project, we learned how to build and connect a modern full-stack application using Next.js and FastAPI. We also gained experience working with APIs, asynchronous requests, frontend state management, Git/GitHub workflows, and deploying a frontend application.

A major learning was that detecting hallucinations is not simply about deciding whether a statement is true or false. Evidence quality, source reliability, confidence, and uncertainty all need to be considered.

Challenges We Faced

One of the biggest challenges was connecting the frontend verification workflow with the backend and ensuring that the returned data matched the structure expected by the UI.

We also had to handle cases where web search or verification services were unavailable, as well as loading states, failed requests, and missing analysis results.

Another challenge was presenting complex verification information in a way that was understandable without overwhelming the user. We addressed this by breaking the results into risk scores, metric cards, claim tables, verdict badges, confidence indicators, and detailed evidence panels.

Future Improvements

In the future, we would like to improve source ranking, add more verification providers, support additional document formats, improve claim-level explanations, and deploy the backend alongside the frontend so that HalluciCheck can be used as a complete production application.

Challenges we ran into

One of the biggest challenges was designing a reliable workflow for breaking a complete AI response into meaningful claims and then verifying those claims against external information.

We also had to handle differences between AI-generated reasoning and information retrieved from web sources, while presenting the results in a way that was simple for users to understand. Connecting the frontend and FastAPI backend and making the analysis flow work smoothly across different components was another important challenge.

Accomplishments that we're proud of

We are proud of turning the idea of AI hallucination detection into a functional end-to-end application. Some of our key accomplishments include:

Building a complete frontend and backend architecture Integrating Groq for AI-powered analysis Integrating Tavily for external evidence retrieval Implementing claim detection and verification Creating a structured risk assessment system Building detailed analysis and claim-result interfaces Adding browser-based analysis history Developing a responsive interface for different screen sizes

What we learned

This project helped us understand how different AI services can be combined with traditional web development to solve a practical problem. We gained experience working with Next.js, TypeScript, FastAPI, API integration, asynchronous workflows, and structured backend services.

More importantly, we learned that building an AI-powered application requires more than simply connecting an LLM to a frontend. Reliable outputs require verification, evidence, careful handling of uncertainty, and a user interface that clearly communicates what the system knows and what it cannot confidently verify.