Synthetic persona engine
Personas carry age, interests, digital fluency, spending habits and mood, so a panel of 10 to 300 behaves like a population rather than one averaged user.
Seoul, Korea · Founded February 2026
Know how people will react before you ship to them.
Scarlatti builds software that tests products with synthetic users. Our first product, Synthia, runs persona-driven test automation against your screens, flows and content, and returns findings in hours instead of weeks.
The problem
User research is the slowest part of a fast loop.
Recruiting participants, scheduling sessions and analyzing results takes weeks, and the cost repeats with every iteration. Early-stage products often have too few users to analyze at all. For habit, health and learning products the delay is worse: behavior change simply takes real time to observe.
The result is that teams ship on instinct, then learn from their first real users, at the expense of those users.
- Weeks to run a single round of moderated testing
- Cold start analytics tools need traffic that does not yet exist
- Months to observe long-term behavior change in real users
Product
Synthia: a research studio staffed by synthetic users.
Upload a screen or point Synthia at a flow. It assembles a panel of AI personas, lets them use the product, and reports where they hesitate, get lost, or leave, broken down by persona and by step.
Multimodal UX understanding
Vision-language models read app screens, web pages and design files, and flag missed buttons, information overload and onboarding confusion.
Agentic browser interaction
Browser agents sign up, search, browse and check out like a person would, so findings come from use rather than from a static review.
Time-compressed simulation
Skip the wait. Run a morning interaction, jump to lunchtime, and observe the follow-up. Weeks of habit formation can be explored in a single run.
Behavioral analytics
Results are organized by persona, funnel step and behavior type, with drop-off likelihood and a plain-language reason attached to each finding.
Built for repetition
Re-run the same panel after every design change. Because personas are versioned, differences between runs reflect your product, not a new sample.
How it works
From screen to findings in three steps.
-
01
Bring a screen or a flow
Upload screenshots or design files, pick a sample scenario, or connect a staging URL for browser-based exploration.
-
02
Compose a panel
Choose 10 to 300 synthetic users by demographic balance, behavior profile and digital fluency. Add the questions you need answered.
-
03
Read the findings
See first reactions, step-level friction and drop-off risk per persona, then run longer scenarios to test whether behavior holds over time.
Traditional user testingWith Synthia
ParticipantsRecruited and scheduledGenerated on demand
TimeReal time: hours to monthsCompressed simulation
Cost per roundRecruiting, moderation, analysisAutomated and repeatable
TimingAfter launch or betaBefore launch
DepthFunctional pass or failBehavior and emotion
Synthetic results are a pre-launch signal for prioritizing what to fix and what to test with real people. They do not replace research with real users.
Where it applies
Anywhere behavior over time matters.
Product & UX
Onboarding optimization, retention analysis, screen-level usability checks before engineering time is spent.
Healthcare & wellness
Habit formation, exercise, sleep and diet programs, where behavior change is the whole product.
Education
Learning persistence, reaction to feedback, drop-off across a course.
Content & media
Reader response to serialized fiction, engagement with video, thumbnail click-through.
Technology
A TypeScript-first stack, built around agents.
Synthia treats a synthetic user as a long-running agent with memory, goals and a clock. The engineering problem is keeping thousands of those agents consistent, observable and cheap to run.
Application
Node.js TypeScript REST / WebSocket APIs Edge delivery
GenAI & agents
LLM orchestration Agentic AI Multimodal models Tool use & browser agents Persona memory
Simulation
Time-compressed scheduler Behavior & emotion models Persona versioning
Data
Data warehouse Event pipelines Behavioral analytics Machine learning
Quality
Test automation Evaluation harnesses Reproducible runs
Intellectual property
A patent application covering time-compressed, persona-based behavior simulation is in preparation.
Responsible by design
Synthia is built on synthetic and demographic-level data. It does not require personal data about real individuals.
Roadmap
Four stages, each shipping on its own.
In development
1 · Screen simulation
Screenshot and design-file input. Output: drop-off likelihood, UX confusion points, reactions by persona.
Next
2 · Browser agents
Synthetic users explore a live staging service: sign-up, search, checkout, onboarding.
Planned
3 · API flow simulation
Drive user flows through APIs for backend-heavy products and repeatable regression runs.
Planned
4 · Content & long-term behavior
Video, serialized content and multi-week behavior change, simulated end to end.
Company
A small team in Seoul, working on one hard problem.
Scarlatti was founded in February 2026 to make pre-launch user research as routine as running a test suite. We are building Synthia in Korea first, where dense mobile usage and demanding consumers make a good proving ground, with a product designed for global teams.
Current status
- Product: Synthia is in active development.
- Funding programs: Our business plan is under review in a Korean government-backed startup competition.
- IP: Patent filing in preparation.
- Legal name
- Scarlatti
- Founded
- February 2026
- Headquarters
- Seoul, Republic of Korea
- Business registration
- 501-15-38911 (Korea)
- Industry
- Software · Artificial intelligence
Contact
Interested in Synthia, partnering, or an early pilot?
We are speaking with product teams in health, commerce, education and content who want to test earlier. Write to us and the founder will reply directly.