Shadow Market
Gemini 4 NYC HackathonShadow Market: a live pitch room with two AI assistants The founder pitches out loud, and two AI assistants listen to the same conversation in real time. Judge's assistant: while the founder is still talking, a team of six Gemini agents researches the idea with live Google Search: market size, existing competitors, startups that failed or succeeded, go-to-market, social virality and sales. A 200-buyer panel scores it on 5 rubrics with the Decisions API (1,000 calibrated answers in about 3 s). When the pitch ends, judges get charts and the 3 sharpest questions, checked against what the founder actually said. Founder's assistant: whatever a judge asks, the founder gets a full spoken answer in about 2 seconds, then an update fact-checked by a live web search of that exact question. Who spoke is decided from the content by the Decisions API, so one open mic works for the whole room. Behind it: an Antigravity Managed Agent builds the MVP in a Google sandbox, AI testers try it in real browsers (recorded), and the agent patches V2 and retests it on held-out buyers. Proof: the launch probability is backtested on 120 real 2026 Kickstarter campaigns (analyst AUC 0.88 vs 0.70 for our buyer swarm). We caught and removed our own data leak, and changed the product when our original thesis lost. Models: Gemini 4 Hydrogen, Neon, Argon, Carbon · Decisions API · Gemini 3.5 transcribe-live · Gemini 3.8 Live · Gemini 3.5 Flash-Lite · Antigravity Managed Agent. Track: Problem Statement 1 (5 Sigma) + Managed Agents.





