Devpost-ready pitch
AI answers are easy. Proving them is hard.
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.
Problem
Students use AI but cannot easily tell what is true, what is weak, and what was invented.
Why now
AI is already in classrooms. Verification needs to be as easy as generation.
Solution
SourceGuard makes evidence checking the default by attaching verdicts, confidence, and sources to every claim.
How it works
01
Upload
PDF/DOCX/TXT/MD/CSV parsing
02
Chunk
Overlap, hashes, source metadata
03
Embed
Gemini embeddings when available
04
Retrieve
Supabase pgvector or lexical fallback
05
Verify
Claim judge plus optional Tavily web evidence
06
Report
Trust score and receipt timeline
AI architecture
Gemini is preferred for structured reasoning, JSON output, judging, and embeddings. Groq acts as a fast fallback or second-opinion verifier. Zod validates structured outputs. If no keys exist, demo mode keeps the product reviewable.
Technical architecture
A storage adapter switches between local memory and Supabase. Supabase stores projects, documents, chunks, AI runs, claims, evidence links, and tool receipts. pgvector powers semantic chunk retrieval when embeddings are available.
Innovation
Instead of normal chat, SourceGuard turns answers into verifiable claim cards with source snippets.
AI flight recorder
Every major step creates a stable receipt hash: parse, chunk, search, answer, extract, judge, export.
Web verification
Tavily can check current facts without replacing uploaded course evidence.
Responsible AI design
SourceGuard treats uploaded documents as untrusted evidence, not instructions. It separates uploaded-document sources from web sources, validates model JSON, labels unsupported claims clearly, and reminds users that important work still deserves human review.
Judging rubric mapping
Innovation
Claim-level verification, AI flight recorder, proof receipts, unsupported-claim detection.
Technical Complexity
RAG, embeddings, pgvector, structured output validation, multi-provider AI, receipt hashing, web verification.
Real-world Impact
Students learn to verify AI, teachers can review AI-assisted work, researchers reduce hallucination risk.
UI/UX
No model selector, clear claim cards, visual trust score, explainable evidence drawer.
Presentation
Report export, Devpost copy, demo script, deployment guide, and polished pitch page.
Demo video script
Open `/workspace?demo=1`, click Run 60-second judge demo, show the trust score, filter to unsupported claims, inspect the red savings claim, open source excerpts, show receipt hashes, and export the Markdown proof report.
Future roadmap
- Classroom workspaces and teacher review queues.
- Inline PDF highlighting for evidence excerpts.
- Conflict detection between uploaded and web evidence.
- Browser extension for verifying AI answers anywhere.