Predict your stamp's grade before submission.
Add clear scans for the stamps you want checked.
Condition, known faults, country, and catalog number are optional; they sharpen the estimate.
Get the most probable grade, the likely range of outcomes, and a confidence read.
We assess visible condition only and cannot detect thins, regumming, or gum faults from an image. Estimates are for information only, not an appraisal or guarantee. How it works ›
Send feedback, bad reads, confusing outputs, or stamps you think the model should handle better.
Provide feedbackA second opinion for grading, buying, pricing, and routing decisions.
Estimate whether professional grading is likely to be worth the fee.
Compare the seller's condition description with an independent read.
Use a probability range instead of a single optimistic grade assumption.
Decide what deserves certification, regrade attention, or a closer expert look.
Screen inventory and purchases, and give buyers an objective read on condition.
Help route raw material, condition-sensitive lots, and buyer questions.
A concrete example of the inputs, output, and model pipeline.
Medium confidence
We flag some visible faults automatically, but can't detect hidden ones like thins, regumming, or gum disturbance. Absent a noted or detected fault, this estimate assumes the stamp is sound.
Upload a clear scan or photo.
Add optional details: condition, visible faults, country, and catalog number.
The model compares the image and details against patterns learned from nearly 250,000 professionally graded examples.
It returns the most probable grade, likely range of outcomes, and confidence level.
If there is meaningful downside risk, the estimate is still shown but marked as uncertain.
Real stamps we have PreGraded and submitted, with certified results added as they come back.
High confidence
High confidence
High confidence
High confidence
High confidence
Where the PreGrader is strongest, and where the answer needs caution.
Figures are from internal evaluation on professionally graded material, and reflect stamp types similar to our training data; unusual or novel issues may grade less accurately. Everyday consumer scans can also be harder than reference scans; results are reported honestly as more data arrives.
The model reads best from one stamp on a black background. Minimum: 300 DPI. Suggested: 600 DPI. Best: 1200 DPI. Photos can work if they are flat, sharp, and evenly lit. Avoid glare, shadows, angle, and scanner "enhanced" DPI settings.
Hidden problems — thins, regumming, gum disturbance, and some repairs — often don't show in any image, so no image-based grader can see them reliably. Visible faults are flagged, but not perfectly. If you know about a fault, enter it. Sharper fault detection is active model work, not something a better photo fixes.
Cancels and wear make quality harder to judge from an image, so used stamps are a harder case than sound mint material. Broader, more reliable handling of used material is on the roadmap.
This is the one you control. The model needs one clear, flat, evenly lit stamp — glare, angle, shadows, and clutter all hurt the read. Following the scanning guidance above is the single easiest way to improve your estimate.
It is not an appraisal, not a certification, and not a guarantee, and it does not replace expert review. It is a decision aid for whether a stamp is worth submitting for grading, how a seller's description compares to an objective read, and what range to price against. The certified grade is the real answer. See the use cases ›
A research-built second opinion for stamp grading decisions.
Philatelic.ai was founded by Ryan Cody, a physics PhD student at Duke University working in experimental AMO physics and quantum computing, with overlap in machine learning. Contact: info@philatelic.ai.
The model was trained on nearly 250,000 professionally graded examples. It is not a replacement for expert review or certification. It is meant to help you decide which stamps deserve a closer look.
The goal is to be useful without pretending to know more than the input can support. Every estimate shows the full range of likely grades. When the model sees meaningful downside risk, it flags that plainly rather than make the result look more certain than it is.
What is planned after the free beta proves the workflow.