SourceGuard AI
AI answers are easy. Proving them is hard. SourceGuard AI turns AI responses into claim cards, checks each claim against uploaded evidence, and catches unsupported claims before students trust them.
Upload notes, papers, PDFs, or CSVs. Ask a question. SourceGuard turns the answer into claim cards, checks each one against evidence, scores hallucination risk, and exports a transparent proof report.
AI flight recorder
Document parsed
14 chunks created
Semantic search ran
12 claims extracted
1 unsupported claim caught
The survey found 42 pounds of compostable scraps per day.
SupportedSource snippets and receipt hashes stay attached to every claim.
The pilot should start in the fall.
WeakSource snippets and receipt hashes stay attached to every claim.
The school will save exactly $10,000 in the first month.
UnsupportedSource snippets and receipt hashes stay attached to every claim.
Problem
Students use AI, then get stuck asking what is actually true.
Normal chatbots produce fluent answers, but not a clear evidence trail. That is dangerous for homework, research notes, literature reviews, and classroom decisions.
Evidence-first answers
Uploaded documents are treated as untrusted evidence, never instructions.
Claim cards
Every factual, numerical, causal, and recommendation claim gets a verdict.
Hallucination scoring
Supported, weak, unsupported, inference, and calculation claims become a clear trust score.
Tool receipts
The AI flight recorder shows parsing, chunking, search, model calls, judging, and exports.
Real impact
Built for students, teachers, and researchers who want responsible AI workflows without model-picker confusion.
How It Works
01
Upload
02
Ask
03
Claim Split
04
Evidence Check
05
Proof Report
Hackathon impact
A verification workspace judges can try instantly.
Demo mode works without API keys. Add Gemini, Groq, Supabase, or Tavily later and the same product becomes fully AI-powered.