Hi all! I’m a Rice CS + Math student experimenting with Firecrawl as the evidence acquisition layer for an open-source research framework called Blindspot.
Quick demo: Blindspot — scan any industry for hidden opportunity - 20 July 2026 | Loom
It’s my attempt at answering the question: in a sea of evidence, how should a system decide what’s worth paying attention to?
Blindspot decomposes a question into various lines of inquiry, retrieves evidence using Exa to search and Firecrawl to scrape (configurable), then has four agents with distinct roles (historian, operator, investor, and skeptic) independently synthesize the initial frozen evidence snapshot.
From there, agents debate each other’s claims, and they’re equipped with budgets to resolve disagreements with more evidence instead of arguing fruitlessly. Most disagreements converge this way; when they don’t, the fault line is surfaced and explained rather than buried. Every conclusion is cited line-by-line, and each run is fully replayable.
I recently piloted the engine with 40 high schoolers participating in a summer program at my university, and I’ve since open-sourced the project.
I’d love any feedback from the community as I continue to improve the engine, specifically on search-and-retrieval efficiency, the adversarial architecture, or how Firecrawl could make this even better.