Focus groups cost $15K and take weeks. What if you could run one in 30 seconds?
YAML-defined personas, multi-LLM support, structured output, real-time cost tracking. Not a replacement for real research — a 1000x cheaper complement for early-stage validation.
pip install synth-panel
The Interesting Problem
User research has an iteration speed problem
Good product decisions need user input. But traditional focus groups cost $5K–15K and take 2–6 weeks to recruit, schedule, moderate, transcribe, and analyze. That timeline means you run 1–2 rounds and hope you asked the right questions. If you didn't, you won't find out until you've already shipped.
The hypothesis behind Synth-Panel is that synthetic personas — LLMs role-playing detailed demographic and psychographic profiles — are useful for early-stage validation. Not as ground truth. Not as a replacement for talking to real humans. But as a fast, cheap way to stress-test your questions, catch obvious blindspots, and iterate on your research instrument before spending $15K on the real thing.
Running the same panel across multiple LLMs (Claude, GPT-4, Gemini) provides a form of triangulation — if all models converge on the same insight, it's worth investigating with real users. If they diverge, you've found where your question is ambiguous or your personas are underspecified.
How It Works
Define personas in YAML. Write your survey. Get structured responses in seconds.
panel.yaml survey.yaml personas: questions: - name: Sarah Chen - "What would you pay?" age: 34 - "What feature matters most?" role: Product Manager follow_up: true income: $120K values: [efficiency, data] \ / \ / v v [Synth-Panel Engine] |-- persona hydration |-- parallel LLM dispatch |-- cost tracking \-- response collection | v results.json ($0.03 total, 4.2s)
Define
Write YAML personas with demographics, psychographics, and behavioral profiles. As rich or minimal as your research requires.
Run
All personas execute in parallel across your choice of LLM. Cost tracked in real-time. Results stream as they complete.
Analyze
Structured JSON output feeds directly into analysis pipelines. No transcription, no manual coding, no ambiguity.
Watch It Run
A 3-persona panel on product pricing preferences. Total cost: $0.03. Total time: 4.2 seconds.
Traditional vs. Synthetic
Not a replacement for real research. A complement that's 1000x cheaper for early-stage validation.
| Traditional | Synth-Panel | |
|---|---|---|
| Cost | $5,000 – $15,000 | $0.01 – $1.00 |
| Time | 2 – 6 weeks | 30 seconds |
| Iterations | 1 – 2 rounds | Unlimited |
| Panel Size | 6 – 12 people | Unlimited personas |
| Output | Transcripts + summary | Structured JSON |
Technical Highlights
The design decisions behind Synth-Panel.
YAML-Defined Personas
Define panelists with demographics, psychographics, and behavioral profiles in plain YAML. Rich or minimal — match the fidelity your research question needs.
Parallel Execution
All personas run simultaneously. A 20-person panel completes in the time it takes for one response. No scheduling, no no-shows, no rescheduling.
Multi-LLM Support
Claude, GPT-4, Gemini, Grok, Llama — run the same panel across different models to compare response patterns and reduce single-model bias. That's triangulation.
Real-Time Cost Tracking
Cost per panelist, per question, per run. A full panel typically costs under $1. Know exactly what each insight costs before you commit to a larger study.
Structured Output
JSON/NDJSON output ready for analysis pipelines. Stream NDJSON for real-time processing or batch JSON for downstream tools. No parsing PDFs or transcripts.
MCP Server Support
Expose panels as Model Context Protocol tools. AI agents can run research panels as part of larger workflows — market research on autopilot.
Plugin System
Extend with custom analyzers, exporters, and persona generators. Build plugins for your specific research methodology or integrate with existing tools.
Structured Instruments
Design surveys with branching logic, follow-up probes, and skip patterns. Likert scales, open-ended, ranking — the same instruments used with real respondents.
Quick Start
Install to first panel in under a minute.
# Install
pip install synth-panel # Run a panel
synth-panel run --personas panel.yaml --survey questions.yaml # Run with a specific model
synth-panel run --personas panel.yaml --survey questions.yaml --model claude-sonnet-4-20250514 # Start MCP server
synth-panel mcp serve Built With
Open source. Free forever.
Install and run a panel in under 60 seconds. Contribute on GitHub.