VisorVolt: 40 images. Any domain.

// The Proof

39 datasets. None of them ours.
None of them failed.

Assets first, then damage. Every box is a real detection on a held-out frame.

Assets

Pass mark set before we ran: under 55% of a fully-labeled model counts as a failure. Zero of 39 hit it — no per-domain tuning, no architecture changes.

Telecom, real detections from a 40-label VisorVolt model — Telecom Tower joints · public benchmark · `joint` 75%*of a hand-labeled model*

Construction, real detections from a 40-label VisorVolt model — Construction Site safety · `person` 96%*of a hand-labeled model*

Thermal, real detections from a 40-label VisorVolt model — Thermal Night imagery · `person` 77%*of a hand-labeled model*

Microscopy, real detections from a 40-label VisorVolt model — Microscopy Plant stomata · `open` 82%*of a hand-labeled model*

Logistics, chassis detection in a container terminal, real output from a 40-label VisorVolt model — Logistics Container terminal · `otr_chassis` 93%*of a hand-labeled model*

Aerial survey, real detections from a 40-label VisorVolt model — Aerial Survey Overhead drone · `bodysurface` 85%*of a hand-labeled model*

Documents, real detections from a 40-label VisorVolt model — Documents Table extraction · `table` 97%*of a hand-labeled model*

Jute sacks detected on a pallet stack by a 40-label VisorVolt model — Bulk Goods Jute sacks · `sack` 83%*of a hand-labeled model*

Every percentage is measured against a model trained on that task’s entire labeled dataset — the annotation project you would otherwise be paying for. Forty images gets you most of the way there.

Proven across the three ways you actually capture

Drone & aerial Oblique and nadir flight imagery — the telecom, aerial-survey and grid tiles above.

Pole & mast, looking down Fixed cameras over a yard or aisle — the container-terminal tile, and the warehouse counts below.

Ground level & handheld Phone, body-cam and vehicle imagery — the construction, thermal and bulk-goods tiles.

Same 40-image recipe in all three. What changes the score is how big the object is in the frame, not what the camera was bolted to — so tell us the camera and the standoff distance and we will tell you where you will land before you buy.

Defects

Finding the asset is the easy half; the value is in the damage — rare, subjective, and expensive for an expert to box. Same recipe, unchanged.

Civil / Bridges defect detection, real output from a 40-label VisorVolt model — Civil / Bridges Structural cracking · `crack` 79%*of a hand-labeled model*

Agriculture defect detection, real output from a 40-label VisorVolt model — Agriculture Crop mildew · `mildew` 80%*of a hand-labeled model*

Industrial defect detection, real output from a 40-label VisorVolt model — Industrial Cable break · `break` 60%*of a hand-labeled model*

Electronics manufacturing, open-circuit defects detected on a held-out inspection frame — Electronics Mfg PCB inspection · `open` 96%*of a hand-labeled model*

Four industries, one system, nothing rebuilt between them. The cable break at 60% is the weakest and it is here on purpose — thin, low-contrast damage on a busy background is the hardest thing we do, and you should see it before you buy.

// What It Does

Detection is the input.
The output is a decision.

A box on an image is not a business outcome. Every detection leaves as a structured record — what it is, where, how confident — in open formats your systems already read.

// Count

Inventory, from footage you already capture

Stock counted per frame, per aisle, per bay. On a fixed camera, averaging a handful of frames cuts count error to 11–13% — without touching the detector.

// Flag

Defects surfaced at a threshold you set

One dial, three modes: cast a wide net for review, run balanced, or flag only what it is certain of. You choose where the false-alarm cost sits — not us.

Blade corrosion detected on a wind turbine

// Export

Into the system that acts on it

Boxes, counts, classes and confidences in open, standard formats — inventory systems, work orders, GIS, compliance logs. Nothing has to move to a new platform.

{ "class": "pallet", "count": 7, "confidence": 0.94, "bbox": [412,688,203,96], "site": "aisle 3 / bay 12", "captured": "2026-09-07T04:18Z" }

// How It Works

From raw imagery
to operational intelligence.

Three phases, one loop. An agent runs it end to end — on your hardware or in your own cloud. We stand it up with you; after that it is yours to run.

40 images per site, once. Everything to the right of that runs on your hardware.

// Phase 01 · Days, not months

Asset Inventory

Upload imagery. Get a structured inventory of the classes you care about.

// Phase 02 · Days per new asset

Auto Expansion

A brand-new asset onboarded in days from 40 images per site — and multiple assets come out of that one pool of 40.

// Phase 03 · Ongoing

Defect Intelligence

Anomalies flagged with confidence scores. Work orders and compliance evidence packages generated autonomously.

Hands-on while we onboard you.Self-serve once it is yours.

Assets, novel assets and defects — detected, self-improved, delivered in open formats. Held-out across 22 public datasets, the loop improved the deployed detector 20 times out of 22, closing a typical 18% of the remaining gap with no further labelling on your side.

// After It Ships

A research scientist
that never sleeps.

A fixed, hand-labelled model starts drifting from field conditions the day it ships, and the usual fix is another labelling round. Yours ships with a research scientist instead. It reads the literature, finds the methods worth trying on your problem, designs the experiment, runs it overnight on your hardware, and keeps only what raises the score — then goes again. One recent experiment lifted accuracy by 4.4 points. Nobody asked it to.

// It reads

Finds the methods worth trying on your problem

New work appears constantly and almost none of it applies to you. The agent reads it against your results and your imagery, and shortlists what is worth an experiment.

// It experiments

Runs it overnight, on your hardware

It builds the experiment, runs it against your held-out data, and measures the result — with no ML engineer of yours in the loop and nothing leaving your network.

Most recent gain: +4.4 points of accuracy, from an experiment nobody scheduled.

// You approve

Nothing ships without this step

See what changed, what it was measured against, and by how much. Then approve or reject.

Autonomous experimentation, human-gated deployment.

That is the half that improves the model. The other half improves the data — the auto-labelling flywheel mines your own archive, verifies what it finds and retrains, with nobody drawing a box. See how the flywheel works → Both are agents, both run inside your environment, and both are designed to hand over — yours to start, pause and audit.

// Why VisorVolt

Every serious vendor trains on your data.
The question is who pays for it.

Every vendor trains on your data. Almost all of them make you label it, one vertical at a time, so each new asset is a fresh annotation project on your budget. Ours generates and verifies its own. The difference is not a benchmark score — it is who pays to expand coverage, and which direction accuracy moves after deployment.

what off-the-shelf scores on this imagery

Proof the domain is hard

Off-the-shelf models score about 8% here, and 0.000 on a class they have never seen. Everyone has to train. The only questions are who pays for the labels and what happens after it ships.

labeling cost to you

Who pays to expand coverage

For label-dependent vendors, every new asset is a project on your budget and your timeline. Ours generates and verifies its own training data — new assets in days, from 40 images per site, several out of the same 40.

cloud required

Beats the cloud. Without the cloud.

Everyone else phones home to the cloud for the hard part — checking what it found. Ours doesn’t. Our on-prem verifier scores F1 0.80 against a leading cloud model’s 0.71, held-out, with nothing leaving your network — and it tells you when it is unsure. The whole pipeline, verification included, runs inside your NERC CIP boundary. Zero external calls, on-prem or in your own cloud.

services retainer to keep it improving

We stand it up. Then it runs without us.

Our engineers deploy forward, on your site, and get it running on your imagery and your hardware. After that both halves — the labelling flywheel and the research scientist — are agents your own team operates: start a run, review what it proposes, promote what you accept. It is designed to run without us, and it improves on its own rather than on a services contract. Hands-on while we onboard you; self-serve once it is yours. See the handover →

9×

fewer false alarms

The number buyers actually feel

Accuracy is not what wastes a crew’s day. False alarms are. On frames containing none of the target assets, the pooled 40-image model emits 11.5 false detections per 100 frames against 104.8 for the three single-asset models it replaces — roughly 9× fewer.

Same category. Different economics.

Any domain. 40 images.
A working detector.

39 datasets. None of them ours.
None of them failed.

Detection is the input.
The output is a decision.

Inventory, from footage you already capture

Defects surfaced at a threshold you set

Into the system that acts on it

From raw imagery
to operational intelligence.

Asset Inventory

Auto Expansion

Defect Intelligence

A research scientist
that never sleeps.

Finds the methods worth trying on your problem

Runs it overnight, on your hardware

Nothing ships without this step

Every serious vendor trains on your data.
The question is who pays for it.

Proof the domain is hard

Who pays to expand coverage

Beats the cloud. Without the cloud.

We stand it up. Then it runs without us.

The number buyers actually feel

Built by someone who
has done this before.

Start with imagery
you're already collecting.

Any domain. 40 images.A working detector.

39 datasets. None of them ours.None of them failed.

Detection is the input.The output is a decision.

Inventory, from footage you already capture

Defects surfaced at a threshold you set

Into the system that acts on it

From raw imageryto operational intelligence.

Asset Inventory

Auto Expansion

Defect Intelligence

A research scientistthat never sleeps.

Finds the methods worth trying on your problem

Runs it overnight, on your hardware

Nothing ships without this step

Every serious vendor trains on your data.The question is who pays for it.

Proof the domain is hard

Who pays to expand coverage

Beats the cloud. Without the cloud.

We stand it up. Then it runs without us.

The number buyers actually feel

Built by someone whohas done this before.

Start with imageryyou're already collecting.

Any domain. 40 images.
A working detector.

39 datasets. None of them ours.
None of them failed.

Detection is the input.
The output is a decision.

From raw imagery
to operational intelligence.

A research scientist
that never sleeps.

Every serious vendor trains on your data.
The question is who pays for it.

Built by someone who
has done this before.

Start with imagery
you're already collecting.