Simulate minds. Predict decisions.
Create choice experiments with AI agents that think, deliberate, and choose — just like your users would.
https://synthchoice.vercel.app/ (Live Demo)
| Landing | Simulation in progress |
|---|---|
 |
 |
| Experiment setup | Results dashboard |
|---|---|
 |
 |
| Conjoint analysis | Choice drivers |
|---|---|
 |
 |
SynthChoice is an AI-powered decision simulation platform for conjoint analysis and choice modeling. Define alternatives, configure agent segments (personality types, demographics, preferences), and run simulations to predict how different user segments would choose.
- Setup experiments — Define feature schemas, add alternatives, configure agent segments (16 MBTI types, 50+ demographics)
- Run simulations — Watch agents evaluate and choose in real time with a visual simulation
- Analyze results — View choice shares, part-worth utilities, willingness to pay, and feature importance
- Node.js 18 or later
- pnpm (recommended) or npm
- Firebase project with Firestore and Anonymous Auth enabled
# Clone the repository
git clone https://github.com/shero4/synthchoice.git
cd synthchoice
# Install dependencies (use pnpm)
pnpm install
# Start the development server
pnpm devOpen http://localhost:3000 in your browser.
- Create a project at Firebase Console
- Enable Firestore Database
- Enable Anonymous Authentication (Authentication → Sign-in methods)
- Copy your config and update
src/lib/firebase/client.js:
const firebaseConfig = {
apiKey: "YOUR_API_KEY",
authDomain: "YOUR_PROJECT.firebaseapp.com",
projectId: "YOUR_PROJECT_ID",
storageBucket: "YOUR_PROJECT.firebasestorage.app",
messagingSenderId: "YOUR_SENDER_ID",
appId: "YOUR_APP_ID",
};| Command | Description |
|---|---|
pnpm dev |
Start development server |
pnpm build |
Build for production |
pnpm start |
Start production server |
pnpm lint |
Run Biome linter |
pnpm format |
Format code with Biome |
- Framework: Next.js 16 (App Router)
- UI: Ant Design
- State: Redux Toolkit
- Backend: Firebase (Firestore, Auth, Storage)
- Visualization: Recharts, Pixi.js
- Language: JavaScript
src/
├── app/ # Next.js App Router
│ ├── layout.js
│ ├── page.js # Home / experiment list
│ └── experiments/
│ ├── new/page.js # Create experiment
│ └── [experimentId]/
│ ├── page.js # Experiment detail
│ ├── run/page.js # Run simulation
│ └── results/[runId]/page.js
├── components/
│ ├── layout/ # App shell
│ ├── experiment/ # Setup (schema, alternatives, segments)
│ ├── runner/ # Simulation runner
│ └── results/ # Charts, shares, conjoint
├── lib/
│ ├── firebase/ # Auth, Firestore, Storage
│ ├── domain/ # Schema, taskgen, simulate, aggregate
│ └── sprites/
└── store/ # Redux slices
A study configuration with:
- Feature schema — Continuous, categorical, or binary features
- Alternatives — Options agents choose between
- Agent plan — Segments with counts and traits (MBTI, location, price sensitivity, etc.)
- Task plan — Tasks per agent, holdouts, repeats
Groups of simulated respondents with shared traits: location, personality type, price sensitivity, risk tolerance, consistency.
- Choice shares — Overall and by segment
- Part-worth utilities — Feature impact on choice probability
- Willingness to pay — Implicit monetary value of features
- Choice drivers — Feature importance from reason codes (LLM agents)
experiments/{experimentId}
├── alternatives/{alternativeId}
├── agents/{agentId}
└── runs/{runId}
├── tasks/{taskId}
├── responses/{responseId}
└── resultsSummary/v1
- LLM-based decision engine (BYO API key)
- Alternative normalization via LLM
- Export results to CSV/JSON
- Advanced analytics (logit models)
- Sprite image generation
- Team collaboration
MIT