Talk vs Facts

Inspiration

Young people learn about money on YouTube. Finance videos make bold claims and price calls, but nobody checks them later. I wanted a simple tool that shows what was said vs. what really happened.

What it does

Paste a finance YouTube link. Talk vs Facts:

• Finds every claim and price prediction in the video
• Checks price calls against real historical prices (Yahoo Finance)
• Checks general claims against trusted sources only (Federal Reserve, IMF, Vanguard, Morningstar and more) via Tavily
• Rates each claim (Holds up, Needs context, Overstated, Data disagrees) with the exact quote, sources and risks

Talk vs Facts rates claims, not people.

How I built it

• NVIDIA Nemotron Nano on Nebius Token Factory reads the transcript and quickly extracts claims and price calls
• Tavily searches an allow-list of trusted finance sources for evidence
• Yahoo Finance provides real price history
• NVIDIA Nemotron Ultra on Nebius Token Factory reasons over the evidence and gives the final verdict
• Supadata for transcripts, Next.js frontend, deployed on Vercel

Challenges

Fairness was the hardest part. Every quote is verified against the real transcript in code, so the AI can't invent quotes. Negative labels need high confidence, and the AI is not allowed to invent deadlines for vague predictions. YouTube also blocked transcript scraping on Vercel, so I switched to a hosted transcript API.

What I learned

Using a small fast model for extraction and a big reasoning model for judging works much better than one model for everything. Guardrails in code matter as much as the prompt.

What's next

Creator track records over many videos, more languages, and a browser extension that shows facts while you watch.

Built With

• css
• github
• javascript
• nebius-token-factory
• next.js
• nvidia-nemotron
• react
• supadata
• tavily
• typescript
• vercel
• yahoo-finance