MarketRank.ai

The challenge

Investors take stock ideas from thousands of creators across YouTube, X, Discord, Substack, and newsletters, but nobody tracks whether those calls actually work. The hard part isn't storing data; it's reliably extracting real buy/sell/hold calls from hours of noisy, rambling video (sponsor reads, conditional calls, week-old recaps) and then scoring them honestly against the market.

What we built

We designed and built an AI pipeline that runs unattended end to end:

• Ingests content from multiple sources: video transcripts, X, newsletters, and Discord.
• Uses AI agents to pull out buy/sell/hold calls, checking every ticker and price against real market data instead of guessing.
• Reviews each extracted call a second time for accuracy, measured against a hand-built benchmark so no change ships unless it improves quality.
• Stores every signal with the quote it came from, then scores its real return against the S&P 500.
• Surfaces creator rankings, dashboards, daily AI-generated stock picks, and alerts.

Impact

• Doubled precision and recall on extracted calls, taking quality from unusable to production-ready so the calls behind every score can be trusted.
• Found and fixed a bug that was silently dropping half of every long video, recovering a large share of the missed calls.
• Cut the cost of processing each video roughly 10×, making full-catalog runs affordable.
• Shipped a production pipeline that runs unattended, with 30,000+ signals scored against the S&P 500.