Founders build great products. They fail at everything after.
Test your launch on a thousand simulated people before you tell one. Crowdsight runs an ad, a price change or a spec change past 1,000 personas grounded in the real market, then an optimiser agent rewrites it until the simulated market response stops improving.
The loop: 1. Ground the category on the live web (Tavily) · 2. 1,000 personas react in independent batches · 3. Code scores per segment and a weighted Launch Score · 4. The optimiser picks one lever (message, price, spec) and writes 3 variants · 5. Variants and the current input race on the same 200 personas · 6. Apply only on a +2 gain, confirm on all 1,000, stop after 3 iterations.
Nebius Token Factory: 6 open model families voice the crowd (Qwen, Gemma, DeepSeek, gpt-oss, MiniMax, Hermes/Llama), a strong model writes the market brief and runs the optimiser, a single-model mono run gives the Diversity baseline, and a vision model reads launch-video frames. Model IDs come from GET /models (npm run models), never hardcoded.
Tavily: competitor and pricing search, customer complaints, and category news become a market brief that every persona reacts inside, so skeptics compare you with real alternatives.
Setup: copy .env.example to .env.local and set NEBIUS_BASE_URL, NEBIUS_API_KEY, TAVILY_API_KEY · npm i · npm run models · npm run dev (port 3100). NEBIUS_MOCK=1 runs the whole UI without keys.
Demo: run once, click Save demo, then open /?demo=1 to replay the recorded run through the same event stream.
Simulated market response, an early signal. Not a forecast.