# Kaleidoscope Overview Kaleidoscope helps insurers build better predictive models using industry-wide claims data. You train your own models on Guidewire Industry Intel data alongside your internal data, without either party exposing record-level data to the other. ## Why Kaleidoscope Claims models hit a predictive ceiling when your internal data runs out. This is most common in: * Thin segments, where you have too few claims to model reliably * New markets or lines of business, where you have little or no history * Low-frequency events, where even a large book has few examples Third-party data products add new *attributes* (columns) to the claims you already have. They can't add claims you don't have. Kaleidoscope adds multi-carrier claims *volume* (rows) to fill those gaps, giving your models more predictive power. Existing Industry Intel models show the kinds of problems where pooled data helps most, such as claims triage, assignment, subrogation, and litigation avoidance, and workers' compensation severity. ## How it works Kaleidoscope uses federated learning: the analytics code moves to where the data lives, instead of the data moving to the code. Only model updates and aggregate results are exchanged. Your engagement involves two workspaces: * **Guidewire workspace** hosts the Industry Intel datasets. Guidewire prepares these datasets for your use case and connects them through a Guidewire task runner. * **Your workspace** is a private environment for your organization. You sign in here, register your own data (if you are using it), and run sessions. Guidewire registers the Industry Intel datasets in the Guidewire workspace, then connects them to your workspace. Connected datasets appear on your *Datasets* page with a *Partner Dataset* label, and you can include them in your training sessions alongside your own data. When you run a session, each dataset is processed by the task runner that hosts it, and the results are combined into a single model. Your connection to an Industry Intel dataset is private. Other carriers can't see that you are connected, and you can't see their connections. Because the Industry Intel data adds rows (more claims) with the same kinds of columns as your data, most Kaleidoscope work uses [Horizontal Federated Learning (HFL)](/train-overview.md). ## Privacy and data protection * Record-level Industry Intel data is never exposed to you or to any party outside Guidewire. * Record-level data from your organization is never exposed to Guidewire or to any party outside your organization. * Other carriers can't see your data or connect to your workspace. ## Where the computation runs How you connect your own data determines where processing happens. Choose the option that fits your IT policy: | Option | Where your data lives | Where compute runs | What you set up | | --- | --- | --- | --- | | No task runner | You only use Industry Intel data | Guidewire | Nothing | | [Hosted task runner](/deployment.md#hosted-task-runner) | Your S3 bucket or Snowflake | Guidewire, with read-only access to your data | A bucket policy or Snowflake access token | | [Edge task runner](/deployment.md#edge-task-runner) | Your AWS, Azure, or GCP cloud, or on-premises | Your environment | IAM roles or service principals, then a task runner | If your policy requires that analytics code runs inside your own environment, use an edge task runner. ## Ways to work * **The workspace UI** is where you manage users, register and validate datasets, and review sessions and results. * **The Python SDK** is where data scientists and actuaries run analysis and modelling, usually from a Jupyter notebook. Guidewire provides notebooks prepared for your use case as attachments on the connected Industry Intel datasets. ## Who does what | Person in your organization | Workspace role | Typical tasks | | --- | --- | --- | | Technical administrator | Administrator | Invites users, assigns roles, registers task runners | | Data engineer | Data Custodian | Registers and validates your datasets | | Data scientist or actuary | Model Builder | Runs analysis and training sessions, reviews results | For details of each role, see [Invite additional users](/quickstart.md#1b-invite-additional-users). With your permission, Guidewire data scientists can also run work on your behalf. They work through your workspace, so their sessions are visible to you. ## What you keep What you can take away from an engagement, such as a trained model or its predictions, depends on your agreement with Guidewire. Confirm this with your Guidewire contact before you start. ## Next steps Follow [Getting Started](/getting-started.md) to sign in and run your first session.