The ground truth your models learn from.

Pre-label with your models, review with your people, and export only what they agree on.

Book a demo →

Pre-label: A predictor labels the task before anyone opens it. Each shape arrives dashed, with how sure the model was.

Groundseed / Traffic
#630.85#2310.71
6kinds of data
Images, video, text, audio, sensors and LLM output, in one tool.
2+people per task
Two to ten per workspace. A label counts when they agree, or when a reviewer settles it.
0unagreed labels exported
Export takes only what was agreed, so you train on ground truth.
1API for all of it
Everything on screen is in the REST API, described in OpenAPI.

01Labelling with AI

Models take the first pass. Connect a detector, a classifier or a segmenter, or type “every helmet” and a vision model outlines each one with your label. Your people only accept, fix or reject.

Outline a whole crowd. A segmentation model traces every person and bag at a busy crossing. One click accepts all 83 outlines, and each one stays editable, point by point.

Photo: a frame of Shibuya Crossing, Tokyo by Basile Morin, CC BY-SA 4.0, outlined by YOLO11 in Groundwork.

Bring any model.

Every workspace runs it first, and its suggestions arrive dashed with their scores.

Connect from
  • Hugging FaceModel hub
  • Weights & BiasesCoreWeave model registry
  • Your own serverAny HTTP endpoint
YOLO11Detect and trackGrounding DINODetect and trackSAM 2SegmentMask2FormerSegmentSigLIPClassify and searchResNetClassify and searchPANNsAudioBERT NERText and LLMsLlamaText and LLMsGeminiText and LLMs
RT-DETRDetect and trackOWLv2Detect and trackMobileSAMSegmentCLIPClassify and searchDINOv2Classify and searchWhisperAudiowav2vec 2.0AudioGLiNERText and LLMsQwenText and LLMs

02Quality

Consensus: two people per task, how many agreed, were settled or held back, and each person's agreement
At least two people agree before anything ships. Ask two people per task, or up to ten. Export takes only what they agreed on, or what a reviewer settled, so what you train on is ground truth.
A task's History: settled by a reviewer, then a trail of accepted, changed and fixed annotations, each with who and when
Every verdict, kept. Accept, reject with a reason, or fix. Each change and every earlier version is kept with who and when, so any label can be audited.
Dataset health: three tasks labelled, two waiting for review, each class labelled against suggested, and object counts and sizes
Know your dataset before it trains. For each class, how many people labelled against how many the model suggested, and what is missing or out of balance. Each bar opens the tasks behind it.

03The team

And the rest of a labelling team’s day.

Catch the wrong label. Clips lays every event of a label side by side. The odd one out is one click from being fixed.
Search by what’s in it. Type “a big cat” and the closest pictures come first. Or filter: has:Bus, objects:0, score<0.5.
Your people, your sign-in. Roles per workspace, reviewers, and single sign-on with OIDC, SAML and SCIM. Every sign-in is logged.

04Data types

One label screen for whatever you label.

Each kind of data is a set of components, mixed per project.

  1. Recordings, sensors, video and GPS on one timeline

    Every source is a lane on the same clock: waveform, spectrogram, a chart per sensor column, video frames and a map that follows the playhead. Label a moment once and it lines up everywhere.

    Used forSound event detection, Activity and fall detection from wearables, Machine faults from vibration

    • Waveform
    • Spectrogram
    • LineChart
    • VideoFrames
    • Map
    • TimeLabels
    The label screen: audio, spectrogram and sensor lanes on one timeline, with labelled moments and a model's suggestions

05Workspaces

A workspace for every project you run.

Run as many labelling projects side by side as you need. Each workspace sets its own data, labels, models, people and review rules. Here are a few teams use it for.

Manufacturing. Mark faults on vibration and temperature on one clock, check inventory rows, and grade the incident reports written about them.

Example workspaces
  • Pump vibration
  • Parts inventory
  • Maintenance documents

06Built by AI

Your data has meaning. Your screen is written by AI.

A project is two files: a model that says what your data means, and a workspace, in the open A2UI format, that says what people see. Change a line and the screen follows. When no component fits, AI writes one to the same contract, previewed on your tasks until someone reviews it.

model: activity
version: 1

sources:
  audio:
    field: audio
    shape: signal
    title: Session audio
  sensors:
    field: sensors
    shape: table
    title: Wrist motion
    index: { column: relative_timestamp, unit: s }

alignments:
  sensors_to_audio:
    type: offset
    from: ${sensors.time}
    to:   ${audio.time}
    offset: 0.4s

annotations:
  moments:
    type: time-range
    on: ${audio.time}
    labels:
      impact: { hotkey: "1" }
      tap:    { hotkey: "2" }

Activity. The model says what your data means: an audio source and a sensor table, how they line up in time, and the moments people mark.

07API

Fits the pipeline you already run.

Everything on screen is in the REST API, described in one OpenAPI file. Generate a typed client for your pipeline’s language with any OpenAPI generator, or hand the file to an AI assistant and have it write one.

PythonTyped client from OpenAPITypeScriptTyped client from OpenAPIGoTyped client from OpenAPIJavaTyped client from OpenAPIC#Typed client from OpenAPIRustTyped client from OpenAPIKotlinTyped client from OpenAPISwiftTyped client from OpenAPIRubyTyped client from OpenAPIPHPTyped client from OpenAPIAny otherAsk an AI to write one
curl -X POST $GS/v1/projects/12/tasks \
  -H "authorization: Bearer $KEY" \
  -d '[{"audio": "s3://field/s-014.m4a"}]'

08Security

Your data stays yours.

Where your files live, who signs in, and what is kept.

Your files stay in your bucket
Recordings and images open from your own storage with your own access. Nothing is copied in.
Your sign-in
Single sign-on with OIDC and SAML, people kept in step with SCIM, and every sign-in logged.
Every change kept
Each label, verdict and earlier version is kept with who and when, so any label can be audited.
Run it where you need
Hosted for you by our team, or on your own servers, where nothing leaves your systems.

OIDCSAMLSCIMYour own bucketAudit historySelf-hosted option

09Pricing

Start with a pilot, then run it your way.

Starting prices in US dollars. We confirm yours after the demo, once we know your data, your team and where Groundseed should run.

PilotStart here$2,500 onceCredited to your first year
Length
Thirty days
People
Your team
Runs on
Our servers
Our team
Sets it up on your data
Start a pilot →
Hosted by us$1,500 a monthStarting price
Length
Monthly
People
Ten, then $49 a month each
Runs on
Our servers, your storage
Our team
Runs, updates and hosts models
Book a demo →
On your servers$20,000 a yearStarting price
Length
Yearly
People
Up to 25
Runs on
Your cloud or data centre
Our team
Sets it up and keeps it current
Book a demo →

Included in every plan

  • Models in the loop
  • Review and consensus
  • History of every change
  • Dataset health
  • Single sign-on with SCIM
  • The REST API

EducationHalf price on any plan for universities, schools and non-profit research labs.Ask about it →