You describe your product idea in a few sentences with enough detail for the AI to understand it. A casting LLM analyzes the description, infers the target market, and generates 4 completely different synthetic customers with unique names, ages, personalities, jobs, habits, and backstories.
Each persona runs on a different LLM model with its own isolated API key and independent memory context, preventing shared reasoning or cross-contamination between agents. Using concurrent API calls, the personas simulate 3 months of product usage in parallel across multiple checkpoints — Week 1, Month 1, and Month 3. Every simulation returns structured JSON containing their experience, satisfaction score, best moment, worst moment, complaints, feature requests, retention likelihood, and willingness to pay.
Because the agents are fully isolated, they naturally disagree. The highest-scoring persona and the lowest-scoring persona are then selected for a 10-turn debate with alternating responses, damage scoring, and live health bars.
A final judge LLM reads the entire transcript along with the raw simulation outputs, retrieves a startup evaluation rubric through a RAG pipeline, and scores the business across multiple dimensions — market potential, retention, defensibility, monetization, and product-market fit. The system also returns confidence scores, comparable real-world startups, and a market analogy explaining what kind of company the idea resembles.
Every persona uses a different LLM model on a different API key — no shared context, no cross-contamination. The disagreement is genuine. One model doesn't know what the others will say.
| Role | Model | Key |
|---|---|---|
| Persona 1 | openai/gpt-oss-20b | ROAST1 |
| Persona 2 | llama-3.1-8b-instant | ROAST2 |
| Persona 3 | meta-llama/llama-4-scout-17b-16e-instruct | ROAST3 |
| Persona 4 | qwen/qwen3-32b | ROAST4 |
| Casting | meta-llama/llama-4-scout-17b-16e-instruct | ROAST_CASTING |
| Judge (RAG) | meta-llama/llama-4-scout-17b-16e-instruct | ROAST_JUDGE |