Kronaxis Panel Studio

1,000 synthetic consumers in 30 seconds. Self-hosted. No human panelists, no recruitment, no incentive payments.

BSL 1.1 · Website · Commercial Licence · DYNAMICS-8 Spec

---

What is Panel Studio?

Traditional consumer research costs £10,000–£50,000 per study, takes weeks to recruit, and pays panelists who lie or rush. Panel Studio replaces the panel with thousands of simulated consumers, each with a unique DYNAMICS-8 personality, a coherent life history, and census-weighted demographics. Submit a product concept, an ad, a policy proposal — every persona responds in character, producing demographically segmented sentiment data in minutes instead of weeks.

It runs entirely on your hardware via a local LLM server. Your stimuli, your sentiment data, and your derived insights stay on your machine.

500 pre-loaded UK personas are included. Start running stimuli immediately after docker-compose up.

How it compares to traditional and AI consumer research

	Panel Studio	Synthetic Users / SyntheticUsers.com	Resemble.ai (synthetic data)	Traditional panel (Quester / Kantar / Nielsen)
Cost per study	Self-hosted: just GPU electricity	£10–50/persona; £500+/study	Subscription (varies)	£10,000–£50,000+
Time to first response	~30 seconds (local LLM)	minutes (cloud)	minutes (cloud)	1–6 weeks (recruitment + fielding)
Personality framework	DYNAMICS-8 (Big Five + HEXACO + 2 digital-age dimensions)	proprietary	proprietary	demographic only
Demographic weighting	Census-weighted via country builders (20 countries)	US-centric	varies	targetable but expensive
Data residency	Fully local; nothing leaves your machine	Cloud only	Cloud only	varies
Reproducibility	Same persona ID + same stimulus = same response	not deterministic	not deterministic	impossible (human variability)
Public falsifiable validation	Yes — see KPM-1 election predictions	No	No	(not the model's claim)
Source available	✓ (BSL 1.1)	✗	✗	✗
Best for	Fast iteration, sensitive stimuli, longitudinal studies	Hosted convenience	Adjacent use case (training data)	Regulatory / publishable studies that demand human panels

If your study needs to be defensible in a regulator's eyes, run a traditional panel. For everything else — concept screening, ad pre-testing, conjoint, longitudinal sentiment tracking — Panel Studio gives you the iteration speed of code with sentiment data that's been publicly validated against real-world outcomes.

Part of the Kronaxis research stack

1. DYNAMICS-8 — the eight-dimension psychographic framework Panel Studio uses to score every persona
2. Panel Studio (this repo) — the engine that simulates 500–65,000 DYNAMICS-tagged personas at a time
3. KPM-1 — pre-registered, hash-verified election predictions (the public proof Panel Studio's outputs map to reality)
4. Kronaxis Router — the LLM proxy that makes running 65,000 simulated personas economically viable

Each piece is independently usable; together they cover the loop from psychographic framework → simulated population → public falsifiable forecast → cost-efficient inference at scale.

Features

• Multi turn conversations with follow-up questions that build on previous responses
• Conjoint analysis to test product attributes and price sensitivity across personality segments
• Focus group synthesis that generates naturalistic group discussions from individual responses
• Panel builder to create new persona panels from demographic specifications
• Scheduled stimuli on cron or interval expressions for longitudinal research
• Export to JSONL, Parquet, and CSV for training data or downstream analysis
• Cross-panel comparison of sentiment across different demographic panels
• Runs locally on a local LLM server with no data leaving your machine

Quick Start

git clone https://github.com/kronaxis/kronaxis-panel-studio.git
cd kronaxis-panel-studio
cp .env.example .env

Set the mandatory values in .env:

echo "TFS_DB_PASSWORD=your_secure_password" >> .env
echo "FLASK_SECRET_KEY=$(python3 -c 'import secrets; print(secrets.token_hex(32))')" >> .env
echo "OLLAMA_MODEL=qwen2.5:3b" >> .env

Then start everything:

docker-compose up -d

The default language model (~2.5 GB) is pulled automatically on first boot. Monitor progress with docker logs kps-ollama -f. Once the model is ready, open http://localhost:8090, navigate to Panels, select "UK Census Panel", and start a conversation.

GPU recommended. The LLM server runs on CPU if no GPU is available, but inference will be significantly slower. A GPU with 6 GB+ VRAM is recommended for interactive use.

API

Panel Studio exposes a REST API for programmatic access. When PANEL_STUDIO_API_KEY is not set (the default), all endpoints are open. Set it in .env to require an X-API-Key header.

For local development, disable the Kronaxis gate (which otherwise requires a free account for exports and panel building):

KRONAXIS_GATE_ENABLED=false

List panels

curl http://localhost:8090/api/panels

Create a conversation and submit a stimulus

# Get the panel ID from the list above.
PANEL_ID="<your-panel-id>"

# Create a conversation.
CONV=$(curl -s http://localhost:8090/api/panels/$PANEL_ID/conversations \
  -H "Content-Type: application/json" \
  -d '{"title": "Product test"}')

CONV_ID=$(echo $CONV | python3 -c "import sys,json; print(json.load(sys.stdin)['id'])")

# Submit a stimulus. Every persona in the panel responds.
curl http://localhost:8090/api/panels/$PANEL_ID/conversations/$CONV_ID/ask \
  -H "Content-Type: application/json" \
  -d '{"stimulus": "A new meal kit service delivers pre-portioned ingredients for 5 meals per week at GBP 45. Would you subscribe? Why or why not?"}'

The /ask endpoint returns immediately with a run_id. Poll /status for progress, or open the web UI to watch responses stream in via SSE.

Check progress

curl http://localhost:8090/api/panels/$PANEL_ID/conversations/$CONV_ID/status

Export results

# JSONL (default), CSV, Parquet, or full JSONL with persona metadata.
curl "http://localhost:8090/api/panels/$PANEL_ID/conversations/$CONV_ID/export?format=jsonl" \
  -o responses.jsonl

curl "http://localhost:8090/api/panels/$PANEL_ID/conversations/$CONV_ID/export?format=csv" \
  -o responses.csv

Architecture

               +------------------+
               |   Panel Studio   |  Flask, port 8090
               |   (app server)   |
               +--------+---------+
                        |
           +------------+------------+
           |                         |
  +--------v---------+     +---------v--------+
  |   PostgreSQL 15  |     |   LLM Server     |
  |   + pgvector     |     |  (local model)   |
  |   port 5432      |     |  port 11434      |
  +---------+--------+     +------------------+
            |
   500 seed personas
   loaded on first boot

Service	Container	Port	Purpose
panel-studio	kps-panel-studio	8090	Flask web application and API
ollama	kps-ollama	11434	Local LLM inference server
db	kps-db	5432	PostgreSQL 15 with pgvector
model-pull	kps-model-pull	--	One-shot init container; pulls the default model

How a stimulus flows: you submit a question to a panel. Panel Studio loads each persona's DYNAMICS-8 profile, life narrative, and conversation memory into a personalised prompt. The LLM generates a response shaped by that persona's personality. Responses are aggregated by age, gender, region, and personality segment. The result is a demographically broken-down sentiment report with individual-level data available for export.

DYNAMICS-8

DYNAMICS-8 is an eight-dimension personality framework built for behavioural simulation. It extends Big Five and HEXACO with two dimensions for digital and economic behaviour: Acuity (digital fluency) and Impulsivity (delay discounting). Each dimension is a continuous float from 0.0 to 1.0 with four granular facets, giving 32 behavioural parameters per persona.

Code	Dimension	What It Predicts
D	Discipline	Comparison shopping, budget adherence, structured decisions
Y	Yielding	Endorsement susceptibility, social proof response, compliance
N	Novelty	Early adoption, brand switching, content diversity
A	Acuity	Digital campaign engagement, platform behaviour, privacy settings
M	Mercuriality	Risk aversion, emotional framing response, crisis behaviour
I	Impulsivity	Purchase speed, notification response, impulse buying
C	Candour	Authenticity preference, luxury vs value positioning
S	Sociability	Word-of-mouth amplification, review behaviour, sharing

The full specification is available at lib/dynamics/DYNAMICS-8.md and kronaxis.co.uk/dynamics. The DYNAMICS-8 framework is also available as a standalone library: github.com/kronaxis/dynamics-8.

Data

The 500 pre-loaded personas are the same ungated dataset available on HuggingFace. Each persona includes full demographics (age, gender, ethnicity, occupation, income, education, location), a DYNAMICS-8 profile, and a life narrative. The distribution is census weighted against ONS 2021 data.

Larger datasets (5,000+ premium personas, 65,000 constituency-level personas, custom countries) are available under commercial licence. See COMMERCIAL_LICENCE.md.

Configuration

All configuration is via environment variables in .env. See .env.example for the full list.

Variable	Required	Default	Purpose
TFS_DB_PASSWORD	Yes	--	PostgreSQL password
FLASK_SECRET_KEY	Yes	--	Flask session secret
OLLAMA_MODEL	No	(see .env.example)	Language model for persona responses
OLLAMA_BUILD_MODEL	No	--	Separate (larger) model for panel building
PANEL_STUDIO_AUTH	No	false	Enable session-based login and multi-tenancy
KRONAXIS_GATE_ENABLED	No	true	Require free Kronaxis account for exports and building

Commercial Use

Kronaxis Panel Studio is source-available under the Business Source Licence 1.1. Free for internal, non-commercial use: research, education, evaluation, and personal projects. Each version converts to Apache 2.0 within 5 years of release.

A commercial licence is required if you:

• Use Panel Studio to generate revenue (directly or indirectly)
• Deploy Panel Studio as part of a production service
• Redistribute Panel Studio or create derivative works

Commercial licences are available and include managed cloud API, premium persona datasets, and dedicated support. Contact [email protected] for pricing.

Patents

Panel Studio and DYNAMICS-8 are protected by UK Patent Application GB 2605150.8: "Consumer Behaviour Simulation System", filed 10 March 2026.

Links

• kronaxis.co.uk
• DYNAMICS-8 specification
• DYNAMICS-8 library
• HuggingFace dataset
• Commercial licence
• [email protected]

---

Built by Jason Duke, Kronaxis Limited