FocusAI | Devpost

Elevator Pitch

FocusAI: Simulate your first 1,000 customers before writing a single line of code. An autonomous market research platform that spins up sceptical AI focus groups to stress-test your product, debate pricing, and predict churn—generating consulting-grade reports in minutes, not months.

Inspiration

Building is easy. Knowing what to build is hard. Most startups die because they build things nobody wants. Traditional validation (user interviews, focus groups) is slow, expensive, and biased by polite friends. We asked: If LLMs can roleplay convincingly, can they simulate the customers who make or break a business? We didn't want a "yes-man" bot; we wanted a simulation of the brutal reality of the market.

What it does

FocusAI is a multi-agent simulation engine that replaces weeks of user research with a 5-minute stress test.

Input:** You enter a product idea, feature, or pricing model.
Simulation:** We spin up a virtual room of 5–15 distinct AI personas (with unique spending power, psychographics, and pain points).
Conflict:** These agents don’t just answer surveys. They debate each other, challenge your pricing, and react to group dynamics in real time.
Verdict:** The system analyses the unstructured chaos using established frameworks (like Van Westendorp) to deliver a clear verdict: Build, Fix, or Kill.

How we built it

We engineered a concurrent multi-agent system using Next.js 16 and Google Gemini 3 Pro. This isn't just a chatbot wrapper; it's a complex orchestration layer.

Concurrent Orchestration: We parallelised agent reasoning so personas think and react simultaneously, creating realistic "interruption-style" group dynamics rather than turn-based robotic chat.
Context Locking: We built a strict persona management layer to prevent "drift". A frugal CFO persona remains frugal even when the group gets excited, ensuring psychographic consistency.
Math <-> Language: We implemented complex economic models (Gabor-Granger & Van Westendorp) directly in TypeScript. The system parses messy, unstructured agent arguments and converts them into type-safe, quantitative price-sensitivity curves.

Challenges we ran into

The "Yes-Man" Problem: Early versions were too polite. We had to engage in aggressive prompt engineering to inject scepticism and conflict. We taught the AI to disagree effectively.
Latency vs. Depth: Generating consulting-grade reports involves heavy computation. We optimised prompt chains and moved analytics to background workers to keep the UI snappy while the heavy lifting happened behind the scenes.
Statistical Translation: Mapping natural language sentiment to rigid mathematical pricing models required multiple iterations of logic validation to ensure the data was actionable, not just hallucinatory.

Accomplishments that make us proud

Real Conflict: The agents genuinely argue. Some call features "bloat"; others call them "essential". It feels like a real room of difficult stakeholders.
End-to-End Pipeline: We successfully bridge the gap between creative AI roleplay and rigid statistical analysis.
Actionable Output: We don't just give you text; we give you an optimal price point, recommended features, etc. (e.g., "$15/month") backed by simulated data.

What's next for FocusAI

Voice Simulation: Audio-native focus groups you can listen to live.
Human-in-the-Loop: Allowing a human moderator to step into the room and guide the AI discussion.
Competitor Injection: Simulating how your personas react when a competitor's product is introduced to the conversation.

Tech Stack

AI: Google Gemini 3 Pro
Frontend/Backend: Next.js 16, TypeScript
Analysis: Custom implementations of Van Westendorp, Conjoint Analysis, Gabor-Granger models and more