Introduction to Kili Technology

Kili Technology is an expert-in-the-loop platform for creating, reviewing and evaluating AI data. It brings domain experts, internal teams and external workforce into a structured review process, and keeps a record of who decided what.

Kili supports images, video, documents (PDF), text, audio and geospatial data. Teams typically use it to:

Evaluate model outputs. Domain experts assess model predictions or generated content against criteria you define, such as accuracy, completeness or tone.
Build golden datasets. Experts create and validate the ground truth used to test or fine-tune models.
Create training datasets. Labelers annotate data at scale, with model pre-annotation and review steps to control quality.

In every project, model predictions, annotations and reviews are stored as distinct label types, each with its author and date. This lets you see exactly what humans changed from the model output. See Kili data format.

How Kili can help you

Set up projects your experts can work in

Define what experts assess with a custom ontology, configured in the UI without code
Use classification, entity detection, relation and transcription jobs to capture structured judgments
Reduce manual work by importing predictions from any model as pre-annotations, OCR, interactive segmentation, video tracking and a configurable labeling environment
Give experts clear guidance with project instructions

Design your labeling and review workflow

Configure multi-step workflows that place experts at specific stages, for example a first review by trained workforce followed by expert review
Route work to the right people with multi-group workflows and asset assignment
Use consensus to have several people label the same asset independently, measure their agreement, and resolve disagreements in review
Choose how work is reviewed: sequential review, or cross-review between people of the same level
Improve over time by raising issues and questions, so labelers and first-level reviewers can correct their work and instructions can be clarified during the project

Measure and trace quality

Track agreement and accuracy with quality metrics, per class and per person
Calibrate against ground truth with honeypot
Monitor progress and quality in project analytics
Find the assets that need attention in the Explore view, or with custom search queries for project admins

Integrate with your stack

Connect AWS, Azure, GCP or S3-compatible storage
Export versioned data in the format your model expects
Automate with the Kili API, Python SDK, plugins and the MCP server
Manage access with predefined roles

Deployment

Kili is available as SaaS or deployed in your own environment. See Hosting and Kili security measures.