Encord

Encord is a data platform for teams training computer vision and multimodal models, with a particular focus on "physical AI": robotics, autonomous vehicles, drones and medical imaging. It brings annotation, dataset curation and model evaluation together so the same data, labels and quality checks follow a model from first training run to production feedback.

Annotation

• One editor for images, native video, 3D LiDAR point clouds with sensor fusion, DICOM and NIfTI medical volumes, audio, documents and time-series data.
• Boxes, polygons, polylines, keypoints, bitmasks and object primitives, with nested ontologies for detailed attributes.
• AI-assisted labeling with Segment Anything, object tracking and interpolation across frames, and import of your own model predictions as pre-labels.
• Customisable workflows chain annotate, review and consensus stages, and data agents can run inside those workflows to pre-label, route or check tasks.

Curation and evaluation

• Index searches large datasets with embeddings and natural language, filters by metadata and quality metrics, and flags outliers and duplicate images so you label the data that matters.
• Model evaluation, RLHF-style pairwise comparison and rubric scoring show where a model fails and route those cases back into training.
• An API and Python SDK, VPC or on-premise deployment, and managed annotation and data-collection services for teams that want labelled data delivered.

Pricing

Plans are Starter (image and video annotation, customisable workflows, self-serve support), Team (adds data agents, performance analytics and model evaluation) and Enterprise (multiple workspaces, SSO, SLA, VPC or on-prem). Encord does not publish prices; DICOM, geospatial, LiDAR and other modalities are add-ons.