Cardinal.

AI-powered grading

The future
of grading.

A frontier computer vision and machine learning team from CMU, applying hierarchical learning models to card grading.

Scroll

Grading has a problem. And consumers know it.

The lottery.

Inconsistent.

Comments questioning grading consistency

No objective standard.

No explainability.

Collectors want a systematic, explainable process

90–100 days. $60+ a card.

$59.99 per card and a 90 to 100 day wait

Fed up.

01 / 06

Cardinal

Our solution.

A hierarchical computer vision deep learning model, developed by the leaders in the field at Carnegie Mellon University.

Live inspection · FRONT · 0%

How it works.

01 — Numeric measurement

We measure the card with pixel-perfect precision: centering, edges and corners. Each measurement becomes a numeric score, backed by annotations on the image.

Centering · front55/45 · 60/40

Corner TL · fray / fill812 / 790

Corner TR · fray / fill846 / 872

Edge L · fray / fill903 / 881

Corner BR · fray / fill731 / 704

TL 812

TR 846

BL 918

BR 731

How it works.

02 — Defect aggregation and rollup

2a Scan the whole card

A low-resolution pass flags areas that need a closer look.

2b Inspect flagged areas

A high-resolution pass identifies and classifies imperfections, with annotated images for review.

whitening · .97

fray · .91

fray · .96

clean · .99

2c Roll up the findings

Defects and measurements combine into consistent subgrades and a numeric grade.

CTR9.0

CNR8.0

EDG8.5

SRF8.5

FINAL8.5

Or use it to support a human grader: the same annotated evidence gives them what they need to make a consistent decision.

The same stack read placentas.

Automated model zooming into an annotated decidual vasculopathy detection on a whole-slide image — Whole-slide pathology

Zooming into an annotated surface scratch on the Griffey scan — Cardinal AI

Before cards, this hierarchy screened placental pathology — finding lesions across billions of pixels per slide, where a miss matters more than a grade. Validated in pathology. Validated in art authentication. Now pointed at cards.

Published at the highest level

Then it learned to check itself.

On a card, the same discipline: every grade traced to something measured, and a system built to know when to doubt itself.

01 Detect → classify

A single placenta holds 3,000–4,000 blood vessels; as few as ten diseased ones put it at high risk of preeclampsia. The hierarchical pipeline detects the vessels, then classifies each one — 99.2% test accuracy.

02 Explain

A second, unsupervised layer maps the cells themselves — red blood cells against trophoblasts — and reduces their organization to one measurable score. No cell annotations. No black box.

03 Self-correct

Healthy and diseased vessels separate completely across training, validation and test sets, so the system can flag its own mistakes. Of 12,833 pipeline predictions it flagged 38; in a blind review of the 33 that were vessels, the study’s perinatal pathologist found every diseased vessel correctly identified.

Figure 7A of the paper: the hierarchical pipeline — a ResNet backbone and feature pyramid propose vessels, then a ConvNext model classifies each one

Figure 2A of the paper: a healthy and a diseased vessel, their stain deconvolution, and the resulting red-blood-cell and trophoblast masks

Figure 6B of the paper: morphology separation score for the training, validation and test sets — the healthy and diseased clusters do not overlap

LeDuc, Cagan et al. — npj Digital Medicine (2026). Read the paper →

99.2%vessel classification, test set

12,833pipeline predictions screened

0cell annotations required

Published at the highest level

The same model learned Rembrandts.

The same model was applied with the same principle — explainable evidence to support a human verdict.

01 The question

Is it a Rembrandt? Built on 1,325 paintings — the master, his pupils, teachers, peers and influences — the model answers with a calibrated probability, not a bare yes or no.

02 The evidence

Unified GradCAM combines 17 convolutional layers of the network, from brushstrokes to global light. Purple argues for Rembrandt; yellow against.

03 The verdict

Contours mark where the evidence is strongest: the face, especially around the eyes — the part experts say the master usually painted himself. Rembrandt probability 99.8%. Highly confident.

A genuine Rembrandt portrait, as analyzed in the study (detail: head and shoulders)

The whole painting under the Unified GradCAM heat map: purple toward Rembrandt, yellow against

Arias-Rosales, Wheelock, Moritz, LeDuc & Cagan — PNAS.

95.8%test accuracy when highly confident

97.9%recall, recall-oriented screening setting

91.5%test accuracy overall, 200 held-out paintings

Published at the highest level

Human–AI collaboration

A layer under the expert —
not a replacement.

Figure 1A of the manuscript: layers of analysis stacked over a painting — connoisseurship, historical documentation, material and chemical analysis, technical imaging, AI pattern-matching — with the human expert at the top

Portrait of a Girl, the painting found in a Maine attic in 2024

Case · the Maine attic find

Found in an attic in 2024.
Sold for $1.41M.
Still debated.

84.7% Rembrandt probability — confident, not highly confident It says how sure it is.

Figure 1B of the manuscript: training-set paintings plotted on the model's probability surface — Rembrandts on the upper plateau, non-Rembrandts below, the decision boundary between — Training set: Rembrandts and non-Rembrandts on either side of the decision boundary. Dotted lines mark the confidence thresholds.

Tested blind against co-author Arthur K. Wheelock Jr. — over forty years a curator at the National Gallery of Art. On 186 paintings the model matched his attributions 86.6% of the time; where both were highly confident, 93.4%.

For grading: the machine measures, flags and explains. Your graders decide.

Academic technology — who we are

Team out of Carnegie Mellon — leading the field.

Jonathan Cagan

Co-inventor · CMU profile

A University Professor and head of Mechanical Engineering at Carnegie Mellon, Cagan has spent decades advancing computational design and AI-driven engineering. As a co-inventor of the technology behind Cardinal, he brings that expertise to hierarchical learning for card grading — turning visual evidence into consistent, explainable assessments.

Philip LeDuc

Co-inventor · CMU profile

The William J. Brown Professor of Mechanical Engineering at Carnegie Mellon and an AAAS Fellow, LeDuc works at the intersection of engineering and biology. As a co-inventor of the technology behind Cardinal, he brings experience applying these methods to complex biological images to card grading — with annotated evidence that supports a human verdict.

Backed by Control Group — a partner bringing capital and hands-on operating support to take our research from the lab to the market.

Learn about Control Group →

For graders · marketplaces · auction houses

The new standard of proof.

Request a technical demo →

Cardinal™

Consumers are finished.

Inconsistent.

Unstandardized.

Expensive.

Slow.

Cardinal.