Is this page only for problems no standard Metrici product can solve?

Mostly, yes. If a standard module already answers the requirement, Metrici will tell you to deploy that instead — it is cheaper and faster. What this page covers is the other case: a detection class that does not exist, a verdict that has to reach a machine, or a workflow that needs software built around it. Several projects that started exactly there later became standard modules, including the toll module built for the Danube ferry.

Do we need cameras, images or video ready before the first conversation?

No, but it changes the quality of the answer. With ten or fifteen minutes of footage from the actual position, an assessment moves from a reasonable guess to a defensible yes or no. If the cameras do not exist yet, tell us where the position would be and what has to be seen from it — we can say whether the task is plausible at that distance and resolution.

Whose data trains the model, and where does it run?

The training data comes from the project: imagery of your parts, packaging, vehicles or environment, annotated against the classes you agreed. Where a customer requires it, the whole system — applications, models, interface and outputs — runs on a server inside their own network, with nothing shared with third parties. That is how the certificate-automation project at Certind was built. Data handling and access arrangements are defined per project.

How long does a special AI project take, and what does it cost?

It depends on how many classes must be trained, how much representative data exists, how strict the acceptance criteria are and how deep the integration goes. One custom class on existing cameras is not the same job as three engines wired into a line controller. That is why the first step is an assessment that produces a written scope and an acceptance test, rather than a number taken from a price list.

Can the detection stop a machine, divert an item or open a barrier?

Yes. On projects Metrici has delivered, detection results drive controller outputs, relays, LED displays, audible alarms and barriers, and results also leave the platform as HTTP POST events or database records for your own systems to act on. One boundary: where a function is a safety function under your machine-safety obligations, it must be engineered and certified to that standard. Video detection is an inspection and process-control tool, not a safety certification.

Do you work with integrators and machine builders, or only with end customers?

Both. Several of the projects linked on this page were delivered with a partner in front: a systems house for the cash-handling deployment, regional integration partners for the port and factory installations. Integrators usually consume detection events over HTTP POST or JSON, read them from a database view, or use controller outputs directly, and keep their own application layer above Metrici's detection.

AI Solutions for Custom Computer Vision Projects

When standard AI is not enough,
we build what you need.

Custom AI development. Real-world applications. Built around your process.

Metrici develops custom computer vision and AI solutions designed around your specific requirements. From specialised object detection and quality inspection to process monitoring, automated counting and system integration, we turn visual information into data your operation can act on.

Custom detection models Quality inspection Counting and process monitoring Integrated with your systems Developed and run on your infrastructure

Indicative comparison for planning. Which column you land in depends on whether the classes exist, whether your data can support training, and what the result has to do.
Question	Standard Metrici module	Standard module + custom training	Fully custom development
What is recognised	Plates, people, vehicles, places, codes	Your objects, as additional classes	Whatever the agreed classes are
Work before go-live	Configure per position	Configure, annotate, train, validate	Define, annotate, train, validate, build outputs
Typical trigger for the enquiry	“We want this measured”	“It must also see our part”	“No product sees it at all”
Integration depth	Metrici interface, displays, controllers	Same, plus event export	Same plus bespoke software, APIs, machine signals
Cost and time shape	Lowest — no development phase	Development phase, then deployment	Development phase per subsystem, then deployment

Indicative comparison for planning. Which column you land in depends on whether the classes exist, whether your data can support training, and what the result has to do.

Question

Standard Metrici module

Standard module + custom training

Fully custom development

What is recognised

Plates, people, vehicles, places, codes

Your objects, as additional classes

Whatever the agreed classes are

Work before go-live

Configure per position

Configure, annotate, train, validate

Define, annotate, train, validate, build outputs

Typical trigger for the enquiry

“We want this measured”

“It must also see our part”

“No product sees it at all”

Integration depth

Metrici interface, displays, controllers

Same, plus event export

Same plus bespoke software, APIs, machine signals

Cost and time shape

Lowest — no development phase

Development phase, then deployment

Development phase per subsystem, then deployment

Cash handling Metrici interface showing custom-object counting events beside a live camera view of cash-handling tables, where cassettes, banknote stacks and labels are inside detection boxes.

Metrici interface showing custom-object counting events beside a live camera view of cash-handling tables, where cassettes, banknote stacks and labels are inside detection boxes.

Every ATM cassette checked before it leaves the dock

Brink's, a global cash-in-transit company, had to verify cassette content at every loading position — a check that depended entirely on the person at the table. Metrici ran a dedicated DNN training covering more than 31 object classes: every Romanian banknote denomination, the cassettes themselves, their lids, the label type used for each denomination, and euro banknotes. Each dock gets its own camera and detection engine, and the alarm sounds at the dock, not in a report the next morning.

31+ AI classes trained18 loading docks per locationCamera and engine per dock

Read the cash-handling case study →

Assembly quality Metrici counting-events screen listing custom-object detections from the sofa production camera, beside a live view of a padded component being handled on the line.

Metrici counting-events screen listing custom-object detections from the sofa production camera, beside a live view of a padded component being handled on the line.

AI checks sofa components before boxing

At the Aramis Group works in Baia Mare, the check before boxing and shipping was a human one. Metrici was trained on the components of a two-person sofa — 10 object classes — and on fiducial markers that tell the system which product is on the line and which operator or supervisor is standing at it: one marker per product, ten for operators, ten for supervisors. If any component is missing, an alarm is raised and a production supervisor must check the product before it moves.

10 component classes2 cameras + 2 engines per line10+ products to be added

Read the furniture assembly case study →

Heavy industry Aerial photograph of a large chemical plant with cooling towers, chimneys, process buildings and overhead conveyor lines.

Catching an overhead cart bearing before it jams

At Solvay Sodi's Devnya soda-ash plant, a jammed bearing on the overhead cart line meant pulling the cart out of service at considerable cost. Metrici reads bearing temperatures from a thermal camera, checks each cart's position on the line, and reads the cart's serial number by OCR — including when the cart is upside down and the number is not where a reader expects it. Every alert carries a photo, a timestamp, a cart number and a temperature, in real time.

1 thermal + 1 IP cameraDedicated AI training1 server

Read the chemical plant case study →

Bespoke module Metrici LPR toll station live view showing two ferry loading cameras, each with a recognised heavy-vehicle plate and direction of travel.

Metrici LPR toll station live view showing two ferry loading cameras, each with a recognised heavy-vehicle plate and direction of travel.

A toll module that did not exist, built for one ferry

The Călărași Danube ferry crossing had to be automated without ever stopping the boats. The recognition existed; the toll logic did not — so Metrici developed the Toll Station module specifically for this project, integrated the operator's tariff categories and cash registers, and installed it live with zero downtime. Once manual ticketing was removed, revenue tripled. The module was later released to all Metrici LPR users.

Revenue tripled13 tariff categories162 vehicles per crossing

Read the Danube ferry case study →

Tell us what you need to detect.

You do not need an AI model or a complete technical specification. Start with the operational problem — the five answers below are enough to begin an assessment.

What to detect

What needs to be detected, counted, inspected or recognised? Describe the condition in the words your operators use.

Where

Where does the process take place, and what already covers that position: a camera, a window, a lamp, nothing at all?

Your data

Do you already have cameras, images or video samples showing the condition — including the bad cases, not only the clean ones?

The action

What should happen the moment the system detects it: an alarm, a flag, a record, a divert, a stop, a message to another system?

Integration

Does the result need to reach another application, machine or controller — and who owns that interface today?

What happens next

Our team can review the scenario and discuss the technical feasibility, potential solution and next steps.

Have a challenge that standard software cannot solve? Let’s discuss it.

Then what happens

You send the description

By email, with text, photos or a phone video of the position — whatever you have. No template and no specification required.

Then what happens

We assess feasibility

Classes, camera position, image quality, limits. You get a yes, a yes-if or a no, with the reason.

Then what happens

You get a written scope

Engines, servers, licences, integrations and the acceptance test the finished system is measured against.

When standard AI is not enough, we build what you need.

Three kinds of people usually land on this page.

Production, quality and process engineers

Integrators, machine builders and software houses

Operations, safety and logistics managers

Your process is unique. Your AI solution should be too.

Inspections that take time and drift

Objects standard software cannot see

Processes that need continuous eyes

Quality tied to one operator’s attention

Data scattered across systems

Detection that has to reach a machine

So we treat it as an engineering project

A custom AI solution, from detection to action.

Custom object detection and classification

Automated quality inspection

Counting and process monitoring

Safety and compliance monitoring

Visual data extraction and recognition

Automated alerts and integrations

What decides the scope

What can we develop?

Quality control and defect detection

Production and process monitoring

Logistics and material handling

Safety and restricted-area monitoring

Agriculture and environmental applications

Specialised recognition projects

Where these applications have landed

Manufacturing and automotive

Chemical and heavy industry

Transport, ports and logistics

Pharma, food and packaging

Cash handling and secure logistics

Commercial and public venues

Mining and aggregates

Document automation

Retail and warehousing

Agriculture and horticulture

Environmental and wildlife

Construction

From a specific problem to a working solution.

Define the requirement

Evaluate the visual data

Develop and train the model

Test and validate

Integrate with the operation

Deploy and improve

What the assessment hands you

What this page does not promise

Real projects. Specific problems. Tailored solutions.

Every ATM cassette checked before it leaves the dock

AI checks sofa components before boxing

Catching an overhead cart bearing before it jams

A toll module that did not exist, built for one ferry

Also documented, also custom

Detection logic for a queued port

Recognition that reports to SAP

A language model inside a registry

How to read these numbers

Built to work with your existing environment.

Video sources

Detection engines

Metrici interface

Hardware and outputs

Software integrations

Deployment options

Detection is the easy half. We also deliver the half that has to run.

A solution built around your requirements

Computer vision and integration in one project

Experience across different applications

A practical route from concept to deployment

Control over your deployment environment

We tell you when software is the wrong answer

Before you send the enquiry.

Tell us what you need to detect.

You send the description

We assess feasibility

You get a written scope

about metrici

When standard AI is not enough,
we build what you need.