One research line
LLM-Based Human Digital Twins for Research and Decision Making began in Feb 2025. Each project took the question further, and the current work is built out as a working prototype, Twinlyses.
- Feb – Oct 2025SozioMimicRepresenting a person from their own qualitative data instead of a generic persona.
- Nov 2025 – Mar 2026MindsetTwinReasoning, memory, and consistency across repeated interactions.
- Prototype · currentTwinlysesConsent-governed twins that show their evidence and confidence, and abstain when evidence runs out.
From static personas to evolving human twins
Traditional personas and customer segments describe a fictional average member of a target group, built from generic, generalized, or point-in-time data. Twinlyses explores a different direction: human digital twins that are generated from voluntarily contributed individual data and evolve as new evidence becomes available.
The aim is not a perfect digital copy of a person. It is to understand when a digital representation can meaningfully approximate a person’s preferences, perspectives, and possible responses, while making its evidence and limitations visible.
Questions it is built to answer
- What makes a digital twin representative of a particular individual?
- How much personal evidence is needed before a twin should answer?
- Can a twin recognize when it has insufficient information?
- How should it handle changing or conflicting evidence?
- How can contributors control when and why their twins are used?
One platform, two sides
For the people behind the twins
Contributor app
A consent-first space where people calibrate their twin, contribute evidence about how they decide, and stay in control of how it is used.
- Calibration across behavioral domains
- A tree of understanding
- Personality map
- Talking to your own twin
- Correcting what the twin learned
- Consent controls
Why people take part
- A mirror, not a profile: the tree of understanding shows what has been learned, and every insight can be confirmed, adjusted, or removed.
- Their own voice in product research, instead of a stereotype built from someone else’s data.
- Control by design: consent decides what the twin may use, and for which purpose.
- Reflection and wellbeing tools that are useful on their own.
For teams asking the questions
Research workspace
Where a research question is put to a pool of twins: choose the audience, design the survey, run it, and read the responses with their evidence and confidence.
- Audience filters and twin-pool selection
- Survey and scenario design
- Generated twin responses
- Confidence and evidence indicators
- Comparison and aggregated insights
Why teams use it
- Answers in minutes: put a question to a cohort of twins before committing to a full study.
- Grounded in individuals: each twin answers from a real person’s consented evidence, not a generic persona.
- Honest results: model confidence, evidence limits, and a suggested next validation step come with every study.
- Comparable across markets: the same question and method across countries and segments, with follow-ups to the same cohort.
From customer voices to product decisions
Product decisions increasingly rely on customer insight from many markets at once, alongside product context and expert judgment. Twinlyses explores how an AI decision platform can bring these together, keep regional perspectives comparable, and turn them into recommendations that state how far they can be trusted.
Many markets, one method
Cohorts are sampled across countries, so the same question can be asked of each market in the same way.
Comparable insight
Segment breakdowns by region, age, and gender make differences visible instead of averaging them away.
Recommendations with limits
Each study closes with what was observed, where the evidence stops, and what to validate next.
Grounded analysis
A data-grounded analyst answers follow-up questions from the study itself, not from general knowledge.
- Generative AI
- LLM agents
- Retrieval-augmented generation
- Natural language processing
- Decision intelligence
- Customer research
How a twin answers
- Voluntary contributionA participant provides selected information such as preferences, values, reflections, survey responses, or reactions to scenarios.
- Consent and purposeThe participant defines which information may be used and for what purpose.
- Personal evidence baseThe contributed information is organized into an individual, context-aware knowledge base.
- Twin-based responseRelevant evidence is retrieved and used to generate a response that attempts to reflect the individual’s perspective.
- Confidence and abstentionIf the evidence is insufficient, outdated, or contradictory, the twin should communicate uncertainty or decline to answer.
- Human review and correctionThe contributor can inspect, correct, update, or remove information connected to their twin.
Beyond the average persona
A convincing AI response is not automatically a faithful human representation.
The research behind it
Twinlyses builds on peer-reviewed work on contextualized survey simulation and AI personas.
Built end to end
From the research question to the working product: concept, architecture, and implementation.
- Research vision and problem formulation
- Platform architecture and full-stack development
- LLM agents, retrieval, and evidence grounding
- Interaction and product design
- Evaluation design
- Consent and governance concepts
What needs to be validated next
A working prototype makes the research testable. It does not establish that simulated responses can replace studies with people.
- Individual fidelityCompare twin responses with held-out responses from consenting participants, including where a generic persona performs just as well.
- Reliable uncertaintyTest whether confidence and abstention reflect missing, contradictory, or outdated evidence.
- Value in a real studyEvaluate where twin-assisted exploration helps research teams formulate questions, and where direct human participation remains essential.
Explore a research conversation
Discuss a study, explore a potential collaboration, or request a live walkthrough. What question would you like to investigate?
Twinlyses is a research prototype, not a commercial service. Simulated outputs are evidence-bounded approximations, not verified predictions of individual behavior.