SourceGuard AI

Inspiration

I built SourceGuard AI because students are using AI more than ever, but it is still difficult to know which parts of an AI answer are actually supported by evidence. A normal chatbot can sound confident even when one sentence is unsupported, exaggerated, or completely made up. I wanted to build a tool that helps people use AI more responsibly by making the proof visible.

What it does

SourceGuard AI is an anti-hallucination workspace that turns an AI response into verifiable claim cards. A user uploads documents, asks a question, and SourceGuard checks each claim against the uploaded evidence. Every claim is labeled as supported, weak, unsupported, or inference. The app also generates a trust score, evidence snippets, and an AI receipt timeline so users can understand why an answer should or should not be trusted.

In the demo, SourceGuard analyzes a school lunch composting proposal. Most claims are supported by the uploaded survey, science notes, and meeting records, but the AI answer includes one unsupported claim saying the program would save exactly $10,000 in the first month. SourceGuard catches it, marks it red, and explains that none of the uploaded documents support that statement.

How we built it

I built SourceGuard AI as a solo project using Next.js, TypeScript, Tailwind CSS, and a server-side verification pipeline. The app parses uploaded files, chunks the text, retrieves relevant evidence, creates or loads an AI answer, extracts individual claims, checks each claim against sources, and generates a proof report.

The project includes a local demo mode so judges can test the full experience instantly without needing API keys. I also designed the architecture to support Gemini for AI generation and embeddings, Groq as an optional fallback, Supabase with pgvector for production evidence search, and Tavily for optional live web verification.

Challenges we ran into

The hardest part was making the project more than just another chatbot. I wanted SourceGuard to show the reasoning process clearly without overwhelming the user. Designing the claim cards, trust score, evidence panel, and receipt timeline took a lot of iteration because the project needed to feel explainable, polished, and easy to judge quickly.

Another challenge was graceful fallback behavior. The app needed to work in demo mode without API keys while still being structured for real AI providers and production database integrations when those keys are added.

Accomplishments that we're proud of

I am proud that SourceGuard AI became more than a normal chatbot. It breaks an AI answer into individual claim cards, checks each claim against evidence, and clearly shows which parts are supported, weak, unsupported, or inference.

I am also proud of the demo because it catches a realistic unsupported claim. The AI says a composting program would save exactly $10,000 in the first month, but none of the uploaded documents prove that. SourceGuard marks it red and explains why.

Another accomplishment is that the app works in a no-key demo mode, so judges can test the main experience instantly. At the same time, the codebase is structured for stronger production features like Gemini, Groq, Supabase pgvector, and Tavily web verification.

What we learned

I learned how to design a more trustworthy AI workflow where the model does not just answer, but also has to prove itself. I also learned more about retrieval, claim extraction, evidence scoring, structured outputs, fallback systems, and building a polished hackathon demo that clearly communicates the core idea.

What's next for SourceGuard AI

Next, I want to expand SourceGuard AI into a full student research assistant with stronger document search, live web verification, teacher review mode, citation export, and support for longer research projects.

Built With

• gemini-api
• groq
• next.js
• node.js
• pgvector
• react
• supabase
• tailwind-css
• tavily-api
• typescript
• vercel
• vitest
• zod