Cortex Forge
Annotation, forged for vision

Cortex Forge is the annotation platform for computer vision teams. Send raw images, video, LiDAR, and point clouds; get back pixel-tight, audited ground truth — labeled by a vetted workforce and cross-checked by a consensus engine that catches the disagreement before it reaches your training set.

  • Audited to a measured IoU on every batch
  • 2D, 3D, video & sensor fusion in one pipeline
  • Your imagery never leaves your single tenant
Overview
Live
$2.4M
Volume
+18.2%
Growth
99.99%
Uptime

The ground truth behind teams shipping vision in the real world

Vantage RoboticsCropline AerialMeridian AutonomyHalcyon Medical ImagingNorthwind Retail VisionStrand Defense Systems

What a labeling partner is supposed to deliver — measured, not promised.

98.7%
Median IoU against the hidden gold set
19h
Median batch turnaround at production scale
11M+
Frames labeled across 2D, 3D & video
0
Datasets reused to train another customer's model
The platform

Every label type,one annotation pipeline.

Stop renting a different tool for boxes, masks, and point clouds. Cortex Forge runs the whole arc — ingest, annotate, review, export — across every modality your perception stack consumes.

Pixel-tight tooling

Bounding boxes, polygons, keypoints, and instance masks with magnetic edge-snapping and a model-assisted brush that paints a full segment from one click — then lets a human tighten the last few pixels where it counts.

3D & sensor fusion

Cuboids in LiDAR point clouds linked to the matching camera frame, so a labeler annotates once and the box stays consistent across every synchronized sensor.

Video that tracks itself

Annotate one keyframe; interpolation carries the object across the clip. The labeler corrects drift instead of redrawing all 900 frames by hand.

Model-in-the-loop

Bring your own model to pre-label each batch. Annotators promote, fix, or reject its guesses — turning raw frames into reviewed ground truth in a fraction of the clicks.

Export to your trainer

Ship straight to COCO, YOLO, Pascal VOC, or a custom schema, versioned per release so the exact dataset that trained a model is always recoverable.

Quality is the product

Ground truth thatearns the name.

A label is only worth what your model learns from it. Cortex Forge treats quality as a pipeline stage — measured, cross-checked, and handled by a vetted workforce on infrastructure you can defend in an audit.

Consensus by default

Critical frames go to multiple annotators in parallel; the engine surfaces every disagreement, routes only the contested labels to a reviewer, and never lets a silent split slip into your training set.

Gold-set scoring

Hidden benchmark frames are seeded into every batch and scored continuously, so annotator accuracy is a live number you can watch — not a quarterly audit you hope went well.

Reviewer escalation

Low-confidence and out-of-distribution frames climb a tiered review ladder automatically, ending with a senior reviewer for the rare cases that decide edge-case behavior.

Screened, trained workforce

Annotators are background-checked, onboarded against your spec, and must pass a domain qualification before a single production frame reaches them — with secure annotation rooms for regulated or classified data.

Single-tenant by design

Your imagery lives in an isolated tenant in the region you choose, encrypted with keys you hold. Cortex Forge staff have no standing access, and nothing is reused to train another customer's model.

Full label provenance

Every box and mask carries who drew it, who reviewed it, the tool version, and the timestamp — the immutable trail a safety case or a regulator expects to see.

Task library

Brief a task. Get labels back this week.

Pre-built annotation workflows tuned per domain — each ships with a labeling spec, a quality rubric, and a workforce already trained on the edge cases that break models.

2D + LiDAR

Autonomous driving

Vehicles, pedestrians, lanes, and drivable surface across camera and point cloud, with occlusion and truncation flags your planner can actually act on.

Segmentation

Medical imaging

Lesion, organ, and instrument masks annotated by clinically trained labelers, dual-read and adjudicated for diagnostic-grade ground truth.

Polygons

Aerial & geospatial

Building footprints, crop rows, and damage assessment over high-resolution satellite and drone imagery, georeferenced on export.

Detection

Retail & inventory

Shelf SKUs, facings, and out-of-stock states labeled at scale for planogram compliance and checkout-free vision.

Anomaly

Manufacturing defects

Scratches, voids, and surface flaws marked on production-line frames — including the rare positives a defect model is always starved for.

Keypoints

Sports & motion

Skeletal pose, ball tracking, and event tagging across broadcast video for analytics and biomechanics models.

From the teams

Perception teams ship faster on Cortex Forge.

We replaced three vendors and an internal labeling team with Cortex Forge. A million-frame dashcam batch that used to take a month now lands in two days, and the consensus engine catches the occlusion splits our own reviewers were missing.

D
Dr. Lena Vásquez
Head of Perception, Meridian Autonomy

Gold-set scoring is the first thing I open every morning. Annotator accuracy per task is right there, so a drift in label quality shows up as a number on a dashboard instead of a mystery in next month's model.

M
Marcus Okonkwo
ML Data Lead, Cropline Aerial

Single-tenant isolation with our own keys was non-negotiable for patient imaging. We got clinically trained labelers, dual-read adjudication, and an audit trail our compliance team signed off on in weeks, not quarters.

P
Priya Raghunathan
Director of AI, Halcyon Medical Imaging
Pricing

Pay per labeled frame, not per seat.

Usage-based pricing that scales with the work, not the headcount. Start on a pilot batch, grow into a continuous pipeline.

Pilot

For teams validating a first dataset.

$0setup
  • Up to 5,000 frames
  • 2D boxes, polygons & keypoints
  • Model-assisted pre-labeling
  • Consensus on critical frames
  • COCO / YOLO export
Most popular

Scale

For teams running a continuous labeling pipeline.

Usage/ frame
  • Unlimited volume
  • 3D, LiDAR & video tracking
  • Dedicated vetted workforce
  • Gold-set scoring & throughput SLAs
  • Customer-managed encryption keys
  • Priority support + onboarding

Enterprise

For regulated, classified, or safety-critical programs.

Custom
  • Single-tenant dedicated region
  • Secure annotation rooms
  • Choice of data residency
  • SSO, SCIM & advanced audit
  • Named annotation solutions architect

Questions before the first batch.

Which annotation types and modalities do you support?

Bounding boxes, polygons, polylines, keypoints, and instance and semantic segmentation in 2D; cuboids and full sensor fusion across LiDAR point clouds and synchronized camera frames in 3D; and frame-by-frame tracking with interpolation for video. Export to COCO, YOLO, Pascal VOC, or a custom schema.

How do you guarantee label quality?

Critical frames are labeled by multiple annotators with a consensus engine that surfaces disagreement, hidden gold-set frames are scored continuously to track per-annotator accuracy, and low-confidence cases escalate through a tiered review ladder. Every batch is delivered with a measured IoU against that gold set, so quality is a number you receive — not a claim you take on faith.

Where does our imagery actually live?

In a single-tenant environment in the cloud region you choose, encrypted at rest with keys you manage. Cortex Forge staff have no standing access, every access is written to an immutable audit log, and your data is never reused to train another customer's model.

Can we use our own model to pre-label?

Yes. Bring your own detector or segmentation model and Cortex Forge runs it on each batch to generate pre-labels. Annotators then promote, correct, or reject its predictions, which cuts the clicks per frame sharply while keeping a human accountable for the final ground truth.

How fast can we start, and how does turnaround scale?

A pilot batch can be briefed and labeled within the week. For production pipelines, median batch turnaround is around nineteen hours and scales with workforce allocation; throughput and accuracy targets are written into an SLA rather than left to chance.

Forge your first dataset.

Send a pilot batch this week and get pixel-tight, audited labels back. Bring the raw footage — we handle the workforce, the tooling, and the quality.