ADX Lab · v2 · Measured on your devicePatent pending
One score for any screen, video, banner or photo.
Drop a screenshot, a recording, a banner or a product photo. In seconds, on your device: one ADX score with its confidence band, where the eye lands first, what is wrong, and the fix that earns the most.
Behaviour · five xD pillars
Attention · modelled, with fixation order
Usability · ten heuristics
Paste the copy for a fourth lens. Nothing leaves your device unless you ask the cloud lens to refine the result: five free a day.
Drop a screenshot or recording here
PNG, JPG, MP4, MOV or WebM. Screens, recordings, banners, posters, packaging or product photos. Up to 6 files; a recording is sampled into 6 frames on your device.
What is it?
Platform
What should the user achieve here? optional
Copy on it optional · paste the words, or leave empty and the text is read from the frames
Attention model
Fast: contrast heuristic, instant. Balanced: multi-scale saliency with interface priors, on your device. Best: trained model, when one is installed on this host. Reading the text runs on your device too; the first use downloads a language model once.
On-device by default: your files never leave this browser. The optional cloud lens sends one downscaled frame and the measurements, only when you press it.
Reading frames…
1 · Read the frames2 · Detect, read the text & measure3 · Predict attention4 · Score, band & fixes
ADX Report
Advanced Digital Experience Score · patent pending · analysed on-device · Rahul Sharma · humen.design/dxs-lab
Audit result
Advanced Digital Experience Score
Your rating first
Expert calibration · blindFrame 1
Numbered pins mark where each problem sits. Hover a pin or a card to see the region.
LowHigh attentionIndicative ADX · 0 to 10
Patent pending—
Cloud lens · optional
Refine with a vision model
Sends one downscaled frame (metadata removed) and the measurements above to the vision model, nothing else, only when you press the button. It returns named attention regions, sharper evidence for each problem, a redesign sketch and a three-sentence summary. Nothing is stored.
5 of 5 free refinements left today
Sending…
The journey, step by step
Flow score · per-frame ADX minus transition penaltiesTrend for this screen
Matched by perceptual hash · remembered on this device onlyWhich is better?
One pick trains the ordering · ten secondsBehaviour signals this screen is likely to trigger
Behaviour-analytics vocabulary · likelihood, not measurementPredictive attention
Saliency estimate from contrast, colour and density · not eye-trackingWhere the first glance lands · frame 1
Ten usability heuristics
UX copy
Text read from the frames
Problems and the probable design solution
Ordered by impact on the scoreQuick wins · this sprint
What I would validate next
How the score is built
ADX, one number Patent pending Three lenses folded into a 0 to 10 score, weighted by what you dropped in. Screens: behaviour 45%, attention 30%, usability heuristics 25%. Banners, print and social: attention 65%, heuristics 35%. Product photos: attention and image quality only; interface heuristics and copy do not apply. Paste the copy and a fourth lens joins at 10 to 25%.
Behaviour lens (5 xD pillars). Five experience-design pillars, 0 to 10: Engagement, Frustration (inverted), Navigation, Technical and Form. Predicted from detected buttons, fields, text, wayfinding and, in recordings, scroll reversals, revisits and blank frames.
Attention lens. A saliency map built from local contrast, colour, edge density and centre bias gives Focus, Clarity, Cognitive demand and Engagement, plus the share of attention landing on the call to action. An estimate in the spirit of predictive eye-tracking models, not a trained one.
Usability lens. The ten classic heuristics, each scored from what is detectable (status bands, back affordances, alignment, labels, error colour, density) and marked not applicable when a capture cannot show it. Pasted copy adds reading grade, CTA wording, jargon and reader focus.
Attention model, three tiers (v2). Fast is the v1 contrast-and-edges map. Balanced is a multi-scale saliency model (intensity, colour-opponency and orientation centre-surround, Itti–Koch class) with interface priors: headlines, filled buttons, faces, images and a top-left reading bias. Best runs a trained model in the browser through ONNX Runtime when one is installed on the host. Every tier also predicts a fixation order, shown as numbered points on the frame.
Calibration (v2). The thin outer arc on the gauge is the current uncertainty band; it narrows as rated screens accumulate per category, and the percentile unlocks at 30. Thumbs, pairwise picks and blind expert ratings feed a weekly re-fit of the lens weights. The cloud lens refines evidence and fixes and names what draws the eye; it never sets a score, and lift is always computed by re-scoring the measurements. Weights v2.0.0 · prior ·
Sources. Behaviour score signals · Predictive attention scores and models · Ten usability heuristics. Trademarks belong to their owners; this lab is an independent, on-device estimate for design triage, not any vendor's measurement.
Why this exists
Finding the problem fast is the easy half. Shipping the fix is the job.
I built this lab to show how I triage a screen: score it, name the behaviour it will trigger, and propose a change a team can build this sprint. The same method runs my design and research practice at scale.
Patent pendingThe ADX method: one on-device score from behaviour, attention and usability lenses.