
Venture Foundry solves a real problem with AI research tools: you can't tell what's true. Ask a chatbot for a market size and you get a number with no source, no method, no date. Venture Foundry fixes this structurally — every module follows Gather → Extract → Synthesize. A live Bright Data SERP search retrieves real results, gpt-4o-mini converts them into structured facts, and only then does gpt-4o reason over those facts. The synthesis model never sees raw web content, only verified evidence — and every claim links back to its source. Four connected modules — Ideation, Market, Cofounder, and Capital — write into one shared startup profile, so by the time a founder reaches investor matching, the system already knows their domain and market positioning. Two details make this more than a search wrapper: Cofounder matching is two-stage and agentic. A plain search only returns articles about cofounder matching, not people. Venture Foundry uses SERP to find LinkedIn profile URLs, then Bright Data's Web Scraper API to turn them into real person records — capped and cached to stay fast. The competitor map reasons per-market. Rather than fixed axes, the model reads every competitor's description and derives the two questions that actually differentiate that specific market, then places the idea and flags the emptiest quadrant as an open question, not a guaranteed opportunity. Security is enforced at the database layer: all six Supabase tables have RLS with zero policies, so the browser's anon key can read and write nothing — verified directly. All real work runs through Edge Functions with the service role, keeping API keys off the client. If Bright Data can't return evidence, the module fails loudly instead of quietly answering from memory. Built on Vite/React 19/TypeScript, Supabase Edge Functions, Bright Data, and the AI/ML API — Venture Foundry shows that AI research can mean every claim is checkable before you rely on it.
10 Aug 2026