AI & Machine Learning · Computer vision

We build systems that read images and find what matters

Not a platform, not a demo — a working system trained on your images and wired into your process.

Reading the structure of a scan or photographwhich page this is, what is on it, what is missing
Finding required elementsa signature, a stamp, a marking, a seal
Comparing against a referencewhat differs from the sample, which part is absent
Counting and locatinghow many objects, and where exactly in the frame
A working modeltrained on your images, with the accuracy it reaches on them stated before the build
5–6 weeksto a specification your team can accept, build against, or take elsewhere
from $30,000Blueprint price, published — credited in full against the build
One image typewe start with a single class of image and the errors that matter on it

This is for you if

A person does this with their eyes, and the volume grows faster than hiringThe work is not hard. There is simply more of it each quarter than there are people to look at it.
You have an archive of images or scans and get nothing out of itYears of photographs, drawings or scanned packs nobody can search, count or check.
The check is visual, but the result has to be explainableSomeone will ask why this was flagged and that was not, and “the model said so” is not an answer.
An off-the-shelf tool did not fit your dataIt worked on the vendor's samples and fell over on your lighting, your angles, your forms.
You have few labelled examples and do not know whether it is enoughNobody answers this. It has an arithmetic answer, below.

Four methods, and the one thing that decides between them

Vision is sold as a capability — the machine can see. It is really a choice between four methods, and that choice is not made on what is technically possible.

MethodWhat it doesWhen to take itCost of error
ClassificationAssigns the whole image to a classFew classes, and they do not overlapA mistake is corrected on the next screen. Cheapest to label, cheapest to fix
DetectionFinds an object and where it sits in the framePosition and count matter, not just presenceA miss costs more than a false positive: you can dismiss a box that should not be there, not one that was never drawn
SegmentationTraces the outline pixel by pixelArea, shape or boundary is the answerBy far the most expensive to label, so the precision has to be worth it
Reference comparisonFinds what differs from a known-good sampleA reference sample exists and is stableA false alarm is cheap, a miss is not — so it is tuned deliberately over-sensitive
The rule underneath the table: the method is not chosen by what is technically possible. It is chosen by which is more expensive — to miss something, or to ask again. Where asking again is cheap, run sensitive and let people dismiss. Where a miss carries a regulator, a recall or a payout, the method changes and so does the labelling budget.

How many labelled images you actually need

The first question every buyer asks, and the one no vendor page answers. There is no single number, but there is arithmetic.

What matters is not how many images you have but how many you have per cell — one class under one set of capture conditions. Lighting, angle, camera, background and format each split the data again, so the same archive gives opposite answers depending on how many cells it must cover:

Five thousand labelled images, split four ways: 3 classes × 1 condition = 3 cells → ~1,670 per cell workable 5 classes × 2 conditions = 10 cells → ~500 per cell usually workable 12 classes × 4 conditions = 48 cells → ~104 per cell thin 20 classes × 6 conditions = 120 cells → ~41 per cell not a dataset Same five thousand images. Four different projects.

A narrow task under controlled lighting works on a few hundred per class; the same accuracy across a dozen classes shot on whatever phone was nearest needs an order of magnitude more, and the difference is not the model. Three things move the arithmetic and we use all three: transfer learning, so the model learns your task rather than learning to see; synthetic and augmented data where the variation is mechanical rather than semantic; and a phased launch with a person in the loop, where each reviewer correction becomes a label. Whether your data is clean and accessible enough to start at all is the broader question, in data readiness for AI.

What we have already built

Client names withheld under NDA. See full case studies →

Aviation · MRO

Reading maintenance packs: structure, signatures, empty cells

Scanned maintenance packs arriving as hundreds of pages of mixed forms. The system classifies each page into one of six document types, detects missing signatures and stamps, finds unfilled checklist cells and verifies date-and-signature pairing, then returns an annotated report before a specialist signs off — advisory by construction. Read the case →

−50% manual review6 document types4 classes of check
Being precise about what this is

This is computer vision applied to documents, and we would rather say so than let it read as something else. The mechanics — reading structure, finding a required element, detecting what is absent, comparing against a reference — carry over to photographs and other media, but the engagement was document control under an aviation rulebook and that is the one we can describe. Where the same mechanic meets a regulated paper process it lives in document intelligence; this page is the vision layer underneath.

What we take on, and what we decline

We buildWe decline
Reading structure and elements on scans and photographs · detecting what is absent · comparing against a reference sample · classifying and routing an incoming stream of images · counting and locating objects in a frame Machine vision on moving equipment and navigation — a different discipline, and a different cluster: Physical AI · biometrics and face recognition · medical diagnosis from images · industrial inspection of manufactured parts, which we have not done and will not present as if we had

The last one is worth stating plainly. Visual quality control of physical products is the largest industrial application of vision and several vendors competing for this page sell it — we have not shipped it. The mechanics above are real; that deployment is not ours to claim.

How the work runs

Three phases. The second is a product you can buy on its own.

PhaseWhat we doWhat we need from youWhat you get
AssessmentImage types, capture conditions, and what counts as an error — agreed with whoever does the checking todayA sample of real images, awkward ones includedAn estimate of achievable accuracy on your data and the labelling volume it implies
DesignMethod chosen against cost of error, metrics fixed, and the point where a human stays in the loopSign-off on acceptance criteria before anything is trainedAn AI Solution Blueprint — from $30,000, credited against the build, yours to implement with anyone
BuildTraining, integration, and monitoring of quality in productionA named owner of the processThe system, plus a written procedure for retraining as conditions drift

What we take off the table

Your images

Where they are stored, who can reach them, and a closed perimeter where that is required — open-weight models on your own hardware, nothing leaving the network. Details in security and compliance.

Your model

Pinned and dated versions, a reproducible build, and an honest position on drift: a vision model degrades as cameras, lighting and forms change, so the retraining procedure ships with it rather than being a later conversation. Operational detail in MLOps.

Your budget

Fixed price against a scope agreed before work starts. No time and materials, and the achievable accuracy stated in the Blueprint rather than discovered in month four.

When we tell you not to do this

There are no labelled images and nowhere to get themIf nobody kept examples and the process cannot generate them, the first project is collection, not modelling.
A rule solves it without a modelA fixed template, a barcode, a checksum, a field in a system of record. If the answer is already structured somewhere, reading it out of a photograph is the expensive route to it.
The accuracy you need is not reachable on that dataBlurred, low-resolution or inconsistently framed images set a ceiling no model clears. Better said in the assessment than after four months of billing.
The camera is on a moving machineNavigation, grasping, autonomous equipment. Different discipline, different cluster — Physical AI.

Frequently asked questions

How many images do we need?

It depends on cells, not on the total. A cell is one class under one set of capture conditions, and what matters is images per cell. Five thousand images across three classes and one condition gives about 1,670 per cell and is workable; the same five thousand across twelve classes and four conditions gives about 104 and is thin. So the answer is a question back: how many classes, and how many conditions?

What accuracy is achievable?

On your images, which is the only place the question can be answered. We measure it on a sample of your real data, awkward cases included, and state the figure before you commit to a build. Treat any number quoted before a vendor has seen your images with suspicion.

How long does it take?

Five to six weeks to a specification with acceptance criteria — the Blueprint, which you can buy on its own. A build on top typically runs four to six months depending on labelling volume and integration depth. First results on your own images come inside the assessment, not at the end.

What if we have very little labelled data?

Transfer learning, synthetic and augmented data, and a phased launch where each reviewer correction becomes a label. None of the three fixes a task with more classes and conditions than the data can ever cover.

Do you fine-tune existing models or train from scratch?

Fine-tune, in almost every case. Training from scratch means teaching a model to see before it learns your task, and there is no payback for that when the general capability already exists and is good.

Can this run in a closed environment?

Yes — open-weight models on your own hardware, images and inference inside your network, no path out where that is required. Whether you should own the hardware is a separate question with an arithmetic answer, usually no on cost alone and yes on residency.

Related practices

Start with one image type

Pick a single class of image and the errors that matter on it. We assess what is achievable on your own data, state the accuracy and the labelling volume, and price the build — inside the Blueprint, from $30,000, credited against it.

Not sure a project is the right first step? Free AI Readiness Score — 3 minutes.

Request a blueprint

Not a quote? Ask a question