NpireDashboard
Visual Perception Model

A new class of AI.

Trained on human visual cognition. Built to predict what a real user will see, understand, and decide — before they act. The first commercially deployed Visual Perception Model.

Because the VPM exists, human on-screen behavior can now be simulated with accuracy — not just clicks and keystrokes, but the perception, hesitation, comprehension, and decision underneath them. That capability did not exist before.

The bet

We’re not building a better usability tool. We’re building the model that the next generation of UX tools will run on.

LLMs took a decade to commoditize. The Visual Perception Model is at the start of the same arc. The first product on it is shipping now.

The thesis

Every other AI in this category processes a screen.
The VPM models the mind in front of the screen.

Why this is a new model class

LLMs were not better search engines.

They were a different kind of object. They predict the next token of language. The architecture, the training data, and the output are categorically different from the information-retrieval systems that came before them — even when both can be pointed at the same task.

The VPM is not a better computer-vision system. It predicts the next act of interpretation a human will make on a given visual surface. The unit of prediction is different. The training data is different. The output is different.

The bottleneck in UX measurement for thirty years has been the same: you can simulate a click, but you cannot simulate a perception. Eye-tracking labs sample a few users at high cost. Computer-vision tools identify shapes. Neither tells you what a real user understands about the screen in front of them. The VPM does.

Inside the model

Three pillars.

Pillar 01

Built on cognitive interference research.

Human visual cognition isn't passive recognition. It runs on parallel automatic and deliberate processing streams, and when a UI's visual hierarchy competes with a user's intent, the streams collide. The measurable result is interference: slowed task completion, hesitation, missed actions, abandonment. The VPM operationalizes that interference at the UI level. Decades of cognitive load theory, attention research, and conflict-monitoring literature live inside the model as a usable substrate — not as cited inspiration, but as the actual mechanism the model runs on.

Pillar 02

Trained on meaning-making, not vision.

Computer vision models recognize objects: "this is a button." That is a fundamentally different task from what the VPM does. The VPM interprets comprehension: "this button is visually buried beneath three competing CTAs, so a real user wouldn't perceive it as the primary action." The training corpus is drawn from human-interpretation data — eye-tracking traces, task-completion telemetry, friction-event tags, comprehension panels, drop-off events — not bounding boxes on pixels. Corpus composition and weighting scheme are proprietary.

Pillar 03

Persona-shaped perception.

A model of perception is only useful if it can be conditioned to a specific human. Each audit run conditions the VPM to a defined persona: same model architecture, persona-conditioned weights for age, expertise, patience, knowledge state, device, locale, and behavioral rules. The output isn't one generic perception of your flow — it's thousands of persona-specific perceptions in parallel, each consistent with how that persona would actually behave. The same run can spin up a frustrated novice on mobile and an expert returning user on desktop without changing the underlying model.

Where it came from

The VPM didn’t start as a product.

It started with mental models for U.S. Navy operators in the Combat Information Centers (CIC) of Ticonderoga-class Aegis cruisers, at the Space and Naval Warfare Systems Command (SPAWAR) in San Diego — programs like TADMUS (Tactical Decision Making Under Stress) and the DEFTT (Decision Evaluation Facility for Tactical Teams) lab, studying how trained operators interpret high-density tactical displays under time pressure, where a missed contact is not an abandoned shopping cart.

That work continued through academic foundations in UCSD’s Department of Cognitive Science and the Human Information Processing program in Communications. Three decades later, the same lens that helped a CIC operator find the right tactical contact on a packed display now helps a real user find the right action on a cluttered checkout page.

The interface changed. The cognition did not.

What the model enables

One model. New capabilities.

First product

Benchmark — competitive UX intelligence at matrix scale.

Personas × tasks × flows is a 3-dimensional matrix human panels can only sample one cell of at a time. An agent fleet driven by the VPM saturates the whole matrix in the same audit window. The first product on the VPM is the proof of concept that the model is real and deployable.

See Benchmark
Built-in capability

A/B testing without live exposure.

Because the VPM is a model of human response, paired-variant experiments can run inside the model in minutes. The losing variant never reaches a real user. Statistical-significance wait times disappear. Per-persona outcome signal is preserved.

Where it goes next

Anywhere a model of human perception is valuable.

Pre-launch design validation. Continuous UX monitoring. Adaptive interfaces that route around persona-specific friction. Accessibility coverage measured against real cognitive load, not checklist conformance. The VPM is platform infrastructure, and Benchmark is the first thing built on it.

The VPM, applied

You get a working behavioral simulator. Feed it any persona, any task, any URL — it hands back a step-by-step play-by-play of what happens: which page the persona abandoned at, where they got lost, what specifically blocked them, and how long each step took.

Want to know more?

For Benchmark customers, partners, investors, and press — Npire is the studio that built the VPM. Reach out and we’ll send you the deeper brief, the methodology disclosure, and the roadmap.