Meduzia
Join our Waitlist

The science behind Meduzia.

I · 01

A 50-year-old problem, solved by simulation.

Every protein is a chain that folds into a shape, and that shape determines what it does: whether it carries oxygen, defends the body or causes disease. For half a century, the only way to know it was to see it in a lab. Until science learned to calculate it.

The nucleosome, with real data from the Protein Data Bank. You can rotate it.26

10300

possible ways to fold a typical protein. Trying them all would take longer than the age of the universe.2

Months

or years of lab work to solve the shape of a single protein.3

200M

structures predicted by AlphaFold. Almost every known protein.4

  1. 1972

    The shape is written in the sequence

    Christian Anfinsen wins the Nobel Prize for showing that a protein's chain of amino acids determines the shape it folds into. And that shape determines what it does in the body. The hard part was still to come: calculating that shape without having to see it.1

  2. 1994

    A blind test, every two years

    CASP is born. Labs solve structures and don't publish them. Predictors send in their answers without seeing them. Then the two are compared. For years, no one came close to what the lab measured.2,3

  3. 2020

    AlphaFold passes the test

    At CASP14, AlphaFold 2 predicts structures with a typical error smaller than the width of an atom, on par with lab methods. The next-best method was almost three times further off.2,3

  4. 2022

    From a hundred thousand to two hundred million

    In half a century, labs had solved about a hundred thousand proteins, each one after months or years of work. The AlphaFold database publishes more than 200 million: almost every known protein.3,4

  5. 2024

    The Nobel, and the next step

    Hassabis and Jumper share the Nobel Prize in Chemistry with David Baker. That same year, AlphaFold 3 no longer predicts a protein on its own: it predicts how it binds with other molecules. With that, Isomorphic Labs designs drugs on the computer before going to the lab.5,6,7
AlphaFold: The making of a scientific breakthrough · Google DeepMind

What made AlphaFold credible was the blind test: it predicted shapes no one had ever seen.

Then it was compared with the lab, and it matched. That's the idea we build on: a replica earns trust by being compared with reality.

I · 02

Medicine already runs trials with patients who don't exist.

Testing a treatment on people is slow and expensive. That's why medicine learned to test on the computer first, with virtual patients built from real patient data: everything gets tested there, and only the best makes it to the trial.

A day for 300 virtual patients

Illustration of the UVA/Padova simulator · drag the chart to pick the time

12:00
AdultsTeensKids
Target range80120160200BreakfastLunchDinner00:0006:0012:0018:0024:00
AdultsTeensKidsIn rangeAfter eatingMeals: 08:00, 13:00 and 20:30

300

virtual patients, accepted by the FDA

300 virtual patients with diabetes

The universities of Virginia and Padua build 300 simulated patients, adults, teenagers and children, with the metabolic variety of the real population. The FDA accepts it in place of animal trials for testing artificial pancreas systems.9,10
What it showed
A regulator accepted a validated model as evidence.
How it's used
As a step before trials in people: you get there with what's already been tested.

Today the law says so too. Since 2022, a drug in the US can reach people with tests done on the computer, and in 2025 the FDA announced it will phase out animal testing.13,14

I · 03

From patients to the people who buy.

With language models, the same idea reached human behavior. If you can build a virtual patient from patient data, you can build a consumer profile from what's known about how people live and decide. Top universities have been testing it since 2023.

  1. 2023

    A well-conditioned language model reproduces how different groups of people answer a survey. It gets a name: silicon samples.15

  2. 2023

    25 agents with memory, plans and reflection live in a simulated town and behave believably. On their own, they invite each other to a party and coordinate to go.16

  3. 2023

    Replicating classic behavioral economics experiments, simulated agents reach results similar to the originals.18

  4. 2024

    Agents built from interviews with 1,052 real people reproduce their answers almost as well as those same people do when they retake the survey. The more history each agent has, the better it answers.17

  5. 2025

    Purchase intent is reproduced better when the profile answers in its own words than when it's asked for a score.20

  6. 2026

    In Nature: a model anticipates the direction of the results of 70 social experiments, including studies that hadn't been published yet.19

Each number belongs to its own study, team and country. Meduzia follows that same path with the consumers of each country.

I · 04

Three fields, one logic.

In proteins, in medicine and at Meduzia, the same thing happens in the same order: build a model, test it, compare it with reality and use it to decide.

01Build a model from what's already known

Proteins

The structures the lab has already solved

Medicine

Data from real patients

Consumers, in Meduzia

What people in a category and region are like, and how they decide

It's the logic that took AlphaFold to the Nobel: first compare with reality, then use it to decide.5

II · 05

How Meduzia puts it to work.

We bring that same logic to your brand: we build synthetic replicas of your consumers with data from your brand and your category, and we ask them your question before you decide.

Thousands of profiles per study. Each one with its own story.

How they buy, what they want, what holds them back and how they decide. So they answer like someone in your market.

8
fields per profile
5
personality traits
4
decision styles
Sample synthetic profile
L

Lucía, 22

Medellín · Design student

SEL estrato 3Impulsive mindset
Buys
Online, guided by discounts and videos
Wants
To feel unique in clothes she doesn't see everywhere
Barrier
Not knowing her real size or the return policy
Openness
Conscientiousness
Extraversion
Agreeableness
Neuroticism

  1. 01Your audience

    You choose who to ask: category, region, ages and socioeconomic level.

  2. 02Your brand's profiles

    We build the profiles with data from your brand and your category: your customers' reviews and comments, your social media, your market's demographics, what you share with us and the studies you've already run, plus live research on your category: trends, reviews, competitors and what's being published across the web. Each profile has its own story: how it shops, what it wants, what holds it back and how it decides.

    From your brand

    • Reviews and comments

      From your store, marketplaces and social media

      ★★★★★ “Beautiful, but order one size up.”

    • Social media

      What your community cares about and how it talks

      Do you ship to Medellín?

    • What you share with us

      Your ideal customer, products, sales and CRM

      CRMSalesIdeal customer

    • Your past studies

      Surveys, focus groups and research

      2025 survey · 1,200 responses

    Your brand's profile

    L

    Lucía

    Impulsive mindset

    Data
    22 · Medellín · SEL estrato 3
    Buys
    Online, guided by discounts and videos
    Wants
    Clothes she doesn't see everywhere
    Barrier
    Not knowing her real size
    Traits

    From your category, live

    • Your market's demographics

      Age, area and socioeconomic level

    • Trends

      What people search for, what's rising and what's falling

      +38% “technical jacket”

    • Competitors

      What they offer and what their customers tell them

      3 brands tracked

    • Research on the web

      Studies, reports and articles about the category

      142 articles and reports read
    All of that becomes each profile's story: how they buy, what they want, what holds them back and how they decide.Illustrative examples

    Your audience · 10,000 profiles

    Profiles
    10,000
    synthetic, not people
    Segments
    5
    by how they shop
    Category
    Fashion
    Women's apparel
    Where
    México
    Ciudad de México · Metro area
    Your audience in the Meduzia app: each dot is a synthetic profile. Sample data.

  3. 03The questions

    As many as you need, each in its own format: pick one, pick several, rank, rate or answer in their own words.

  4. 04The answers

    Each profile answers on its own, in the first person and with its reasons. It says what it thinks: whether it wants it, whether it's unsure or whether it's not interested.

  5. 05The results

    One answer per profile, cross-tabulated by segment like a database: who wants it, what holds them back and in what words they say it.

Meduzia's profiles are synthetic: AI generates them from the market you choose, as the ICC/ESOMAR Code requires.25

This method, applied to your next decision.

We build the profiles with data from your brand and your category.

Join our Waitlist
II · 06

Nobody answers the same way twice.

Give the same people the same survey two weeks later and the percentages shift. That's the yardstick science uses to measure a replica: it should drift from the real survey as much as the survey drifts from itself.17

Which of these two capsules would you buy?

Illustrative example
First roundComparison
Technical capsule-2 points
44%
42%
Wool capsule+3 points
33%
36%
Neither-1 point
23%
22%

3

out of every 100 answers would have to switch options for the two to match.

04812

The band is how much people shift among themselves: from 3 to 5 points. A replication should land inside it.

With 400 people, chance alone moves it by about 2.8 points. That's discounted when measuring.

1002004008001,5003,000

How a replica is validated

  1. 01Surveys the model couldn't have seen

    We only use studies published after the model's cutoff date. If the answer was already on the internet when it was trained, the test doesn't count.

  2. 02Written down before measuring

    We record, with a date, what we'll measure and how, before running anything. That way the result can't be adjusted afterward.

  3. 03Blind

    A third party holds the people's answers. We run the question without seeing them.

  4. 04The distance, question by question

    We measure how many answers out of every 100 would have to move for Meduzia to match the real survey.

  5. 05Against the repeated survey and against something simple

    The benchmark is how much the survey itself moves when it's repeated, plus a simple prediction, like the category average, for comparison.

  6. 06Everything published

    The unadjusted number, how many questions and studies, the model and the version, with a date.

III · 07

In the real world.

Brands use Meduzia to decide before they spend: on production, on media or on a field study.

  • Sample questionIllustrative

    Which of these two capsule collections would you buy this winter?

    Technical capsule46%
    Wool capsule31%
    Neither23%

    The technical one wins overall and among online shoppers. Among price-conscious shoppers, it's a tie.

III · 08

Where it's headed.

The science that changed medicine is only beginning to reach brand decisions. This is what's coming, and what we're building.

  • 01Profiles that update themselves

    Every new review, trend or competitor move is added to your brand's profiles, so they answer like your market does today.

  • 02Replicas that are checked all the time

    Every new study is compared against real surveys, and the system learns from each comparison.19

  • 03Every result with its confidence level

    Measured against previous replications, so you know how much weight each answer should carry in the decision.

  • 04The same logic that already changed medicine

    Just as drugs are now tested on the computer before they're tried for real, brands will test every decision before launching it.13,14

  1. 1Anfinsen, C. B. Nobel Prize in Chemistry 1972NobelPrize.org ↗
  2. 2AlphaFold: a solution to a 50-year-old grand challenge in biology (2020)Google DeepMind ↗
  3. 3Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold (2021)Nature 596, 583-589 ↗
  4. 4AlphaFold reveals the structure of the protein universe (2022)Google DeepMind and EMBL-EBI ↗
  5. 5Nobel Prize in Chemistry 2024: Baker, Hassabis and JumperNobelPrize.org ↗
  6. 6Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3 (2024)Nature 630, 493-500 ↗
  7. 7The Isomorphic Labs Drug Design Engine (2026)Isomorphic Labs, technical report ↗
  8. 8Cheng, J. et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense (2023)Science 381, eadg7492 ↗
  9. 9Kovatchev, B. P. et al. In silico preclinical trials: a proof of concept in closed-loop control of type 1 diabetes (2009)J Diabetes Sci Technol 3(1) ↗
  10. 10Cobelli, C. and Kovatchev, B. Developing the UVA/Padova Type 1 Diabetes Simulator (2023)J Diabetes Sci Technol ↗
  11. 11Badano, A. et al. Evaluation of Digital Breast Tomosynthesis... Using an In Silico Imaging Trial (2018)JAMA Network Open 1(7) ↗
  12. 12Qualification opinion for Prognostic Covariate Adjustment (PROCOVA) (2022)European Medicines Agency ↗
  13. 13FDA Modernization Act 2.0 (2022): nonclinical tests, including in silicoSummary in PMC ↗
  14. 14FDA announces plan to phase out animal testing requirement (2025)FDA ↗
  15. 15Argyle, L. P. et al. Out of One, Many: Using Language Models to Simulate Human Samples (2023)Political Analysis 31(3) ↗
  16. 16Park, J. S. et al. Generative Agents: Interactive Simulacra of Human Behavior (2023)UIST '23, ACM ↗
  17. 17Park, J. S. et al. Generative Agent Simulations of 1,000 People (2024, revised in 2026)arXiv 2411.10109 ↗
  18. 18Horton, J. J. Large Language Models as Simulated Economic Agents (2023; later with Filippas and Manning)NBER Working Paper 31122 ↗
  19. 19Ashokkumar, A., Hewitt, L., Ghezae, I. and Willer, R. Large language models can predict the results of social science experiments (2026)Nature 656, 115-122 ↗
  20. 20Maier, B. F. et al. LLMs reproduce human purchase intent via semantic similarity elicitation (2025)arXiv 2510.08338 ↗
  21. 21Bisbee, J. et al. Synthetic Replacements for Human Survey Data? The Perils of Large Language Models (2024)Political Analysis 32(4) ↗
  22. 22Wang, A., Morgenstern, J. and Dickerson, J. P. LLMs that replace human participants can harmfully misportray and flatten identity groups (2025)Nature Machine Intelligence 7(3) ↗
  23. 23Brand, J., Israeli, A. and Ngwe, D. Using LLMs for Market Research (2026 revision)Harvard Business School, Working Paper 23-062 ↗
  24. 24Responsible AI Integration in Survey Research (2026)AAPOR ↗
  25. 25ICC/ESOMAR International Code (2025)ICC and ESOMAR ↗
  26. 26Davey, C. A. et al. Solvent mediated interactions in the structure of the nucleosome core particle at 1.9 Å resolution (2002). The 3D structure in the proteins chapterProtein Data Bank, 1KX5 ↗
  27. 27AlphaFold: The making of a scientific breakthrough (2020), videoGoogle DeepMind, YouTube ↗
  28. 28The Thinking Game (2024), documentaryGoogle DeepMind, YouTube ↗
The real test

Put us to the test with your own study.

Bring us a survey you've already run. We replicate it blind and you do the comparing: the best way to know how much you can trust Meduzia before you hire it.

  1. 01

    Send us a study you've already run

    The questionnaire and who you ran it with. You keep the answers.

  2. 02

    We run it blind

    Your questions, put to synthetic profiles of that same audience, without seeing your results.

  3. 03

    You compare, question by question

    Your real answers next to Meduzia's, with the distance for each one.

Bring a studyJoin our Waitlist
Your study vs. MeduziaExample
Option A
Option B
Neither

4 out of every 100 answers apart