Skip to main content

​
General

Semilattice tests how your users would react to product decisions before you ship them. It uses real human data to predict how specific audiences would answer questions they’ve never been asked.
A user model represents a specific group of people: UK adults interested in consumer finance, software buyers, developers. It’s built from real human data and predicts how that group would answer new questions. You can use our 50+ ready-made user models or work with us to create custom ones from your own data. In the API documentation, user models are referred to as “populations.”
Every user model is tested using leave-one-out cross-validation: each question in the seed data is temporarily removed and predicted as if it were new, then compared to the real answer. Our UK Consumer Finance user models score 90% accuracy on this measure. You can see the accuracy score for every user model in the dashboard.
Yes. We’d love to work with you to build a custom user model for your users, tested for accuracy using cross-validation. All your data and the model itself remains yours. Get in touch →
Short version: we use it to build and operate your audience model, and nothing else. Your data is never used to train other customers’ models, never accessible to other customers, and never used to train foundation models. We do use aggregate performance metrics (accuracy scores, test results) to improve the methodology of how we build models in general — that is not the same as training on your content. See Security and Privacy for the full detail.
Currently single-choice questions, including scales. Anything you’d put in a survey: feature preferences, pricing sensitivity, competitive perception, behavioural intent. The model works best with questions that have distinct answer options.
Traditional research takes weeks and costs thousands. Semilattice returns predictions in seconds. It doesn’t replace deep qualitative research. It fills the gap between gut instinct and a full study, giving you directional signal before you’ve committed to a direction.
Semilattice is free to use right now. No credit card, no contracts. We’re building in public and want early users shaping what gets built.
Book a call and we’ll walk you through it. Or get in touch if you’d like API access.

​
Developers

Populations are what the API calls user models. A population is a model that represents a specific group of people. When you call the API with a population_id, you’re telling Semilattice which group to predict for. Populations are built from real human data (seed data) that determines prediction accuracy and subject coverage. You can browse available populations in the dashboard or list them via the API. Learn more →
Predictions are how you ask your user model a question via the API. You send a question with answer options to a population, and get back a predicted distribution showing how the group would respond. Predictions return population-level outputs (aggregated percentages), not individual responses. You can run single predictions or group them into batches for a project. Create a prediction →
Tests measure how accurately a population model predicts answers to questions where you already know the real answer. You provide ground truth data (real survey responses) and Semilattice compares the prediction against reality. This is different from the built-in cross-validation, which tests the model against its own seed data. Tests tell you how the model performs on your specific questions. Create a test →
The simulation engine takes a population model, a question, and answer options. It uses large language models to predict how the population would distribute their answers. The engine produces stable results: the same question returns similar distributions across runs. Watch the explainer →
Choose a population, send a question with answer options, and poll for results. Predictions typically return in ~20 seconds. Available via Python SDK, Node.js SDK, or REST API. Python quickstart →
The Semilattice MCP server lets AI assistants make predictions directly. Connect it to Claude, ChatGPT, Cursor, or any MCP-compatible tool. Your AI can ask your user model a question mid-conversation. Set it up →