MTCVision AI · a company of MTC Group SpA

Your cameras already saw it. Nobody was watching.

MTCVision AI puts AI detection on the IP cameras you already have. Unsafe conditions become alarms with the frame that proves them — while there is still time to act, not during the incident review.

Book a site assessment See how it works

No cost and no commitment. You keep the technical report either way.

CAM-04 · ASSEMBLY LINE1920×1080 · 29.97 FPSMDL: BASE-PPE-02

Zone 4 · assembly line--:--:--

Person in zoneMOD-01
PPE missingMOD-02
Vehicle proximityMOD-03
Fire / smokeMOD-04
Abandoned objectMOD-05
Capacity exceededMOD-06

Demonstration sequence. Synthetic data, not a live site.

What it runs on

6: base modules, running in hours
26: modules possible in full-scope projects
ONVIF: existing cameras, from 2MP at 15 FPS
Edge / cloud: or hybrid — your call, not ours

The problem

A recording is evidence. It is not a warning.

Most sites already have the cameras. What they do not have is anyone able to watch every frame of every stream, every shift.

ARecorded, not prevented
Footage gets reviewed after the incident, when the cost has already been paid. The camera did its job. The system around it did not.
BNobody watches forty screens
An operator in front of a video wall misses most of what crosses it. Attention is the constraint, and adding cameras makes it worse.
CThe same zone, again
Without event data, a recurring hazard stays invisible until it repeats. You cannot reduce what you never counted.

Why this is different

Generic detection knows what a helmet looks like. It does not know your plant.

Generic CCTV analytics

One pre-trained model, shipped identically to every site.
Accuracy falls under your lighting, angles, dust, and weather.
Often requires replacing cameras for "AI-ready" hardware.
Detection is the product. What happens next is your problem.

MTCVision AI

Base modules run in hours on the cameras already installed.
The Advanced DL model is trained on footage from your site.
It holds accuracy under the conditions your plant actually has.
Detection, alerting, evidence, review, and reporting as one service.

A generic product cannot make this claim, for a simple reason: it does not have your footage.

Base modules

Six conditions, live in hours.

Each module is a named condition that trips an alarm. Every alert carries the frame, the timestamp, the camera, and the criticality.

MOD-01Person in zone
Entry, presence, line crossing, and person-down detection.
MOD-02PPE use
Helmet, vest, glasses, and any other item you define as required.
MOD-03Vehicles
Detection, flow, presence, and coexistence with pedestrians.
MOD-04Fire and smoke
Early signs of combustion, before a detector would register it.
MOD-05Abandoned objects
Bags, boxes, pallets, and anything that is not standard for the area.
MOD-06Capacity and counting
People counting, area occupancy, and corridor blockage.

Up to 26 modules are possible in full-scope projects, given the infrastructure, time, and team.

Two engines

Coverage on day one. Differentiation by year two.

They are sold together because they solve different halves of the same problem.

Base module

Pre-trained · cloud SaaS · no local GPU

Onboarding in hours, not weeks
No new hardware required
Alerts without development work
Predictable monthly cost
Support included in the plan
Suits perimeter, access control, and HSEC

Advanced DL module

Trained on your data · edge deployment

Models trained on your own footage
Holds accuracy under real plant conditions
Quality control on production lines
Deep integration with SCADA and MES
Exportable to edge without latency
The technical difference competitors cannot copy

The platform

Every alert lands somewhere accountable.

Alerts, evidence, zones, and reporting in one console — so an event has an owner, a timestamp, and a record that it was handled.

Product interface in development. The screens below are the design, rendered live in this page — not a running deployment.

Real-time detection across every stream

Each camera reports its own state: model in use, frame rate, and whether a condition is currently tripped. Critical events surface first.

CAM-01Loading bayBASE-ZONE-01
CAM-04Assembly lineBASE-PPE-02
CAM-07Forklift pathBASE-VEH-01
CAM-09Main gateBASE-ZONE-01

Reference footage

What detection looks like in a real plant.

MTC has not published its own demonstration reel yet. Until it does, the clip below is third-party reference footage showing the class of detection the base modules perform.

Reference video: Construction Worker PPE Detection System Using YOLO v11 — ApyCoderWatch on YouTube

Placeholder. Replace with MTC footage from a client site once permission is in hand.

Governed AI

It improves every month. It never changes itself.

You cannot read our source code. You can read every change we made to your model, and approve it.

01
Detect
The system flags events in production, including the ones it gets wrong.
02
Review
MTC reviews false positives and false negatives against what actually happened.
03
Label
Real cases from your site are labelled and added to the training set.
04
Release
A validated version ships — versioned, documented, and authorized before it runs.

Nothing is retrained silently. Every model version is a record you can audit, which is what makes this an operational service rather than software you were sold.

Security and IT/OT

Your IT lead will block this if it is not credible. It should.

Connecting analytics to plant cameras means touching the operational network. These are the controls that make that defensible.

SEC-01Network segmentation
Physical and logical isolation between IT and OT on the Purdue model, with dedicated VLANs for cameras.
SEC-02Encrypted transmission
AES-256 end to end. RTSP streams are encrypted in transit, never carried in the clear.
SEC-03Controlled remote access
VPN with MFA and hardware token requirement. No standing access, no shared credentials.
SEC-04Immutable event log
System actions, access, and configuration changes are recorded and cannot be edited after the fact.
SEC-05Protected evidence
Incident clips stored segregated, with SHA-256 integrity verification and RAID redundancy.
SEC-06Roles and retention
Admin, operator, and auditor roles. Retention windows are a policy you set, not a default we chose.

Worker privacy

The first objection comes from HR, not IT.

Cameras that watch people are regulated, and your works council will ask about this before your security team does. We would rather answer it on this page than in the third meeting.

Detection is configured around zones and conditions, not individuals. The system reports that a condition occurred, at a place, at a time, with the frame that proves it. It is not built to follow a named person through a shift, and we will not configure it to.

What is stored, for how long, and who is allowed to open it are settings you own. We put them in writing before the first camera is connected.

Chilean rules on workplace video surveillance should be confirmed with your own counsel. We build to the policy you set, and we will help you write it down.

How it works

Four phases. The first one is free and you keep the report.

011–2 weeks
Site assessment
Camera inventory, NVR, network, RTSP access, and the zones that actually matter.
022 weeks
Design
Architecture, KPIs, alert rules, and the technical plan you sign off on.
03scoped from the survey
Implementation
Server, models, zones, integration, and acceptance testing.
04monthly
Operations
Support, tuning, event review, model releases, and the performance report.

What each side brings

We are specific about this on purpose.

Projects like this fail on access and ownership, not on models. Here is exactly what we need and exactly what we do.

You provide

PoE network or compatible switch, with a dedicated camera VLAN
ONVIF IP cameras, from 2MP at 15 FPS, and RTSP credentials
Physical access to installation points
SCADA / OPC-UA / PI credentials where integration applies
Historical footage or images of the real environment for training
A named IT/OT counterpart who can approve network access
Room and stable power for the edge server

MTC provides

Technical diagnosis of the existing infrastructure, at no initial cost
Solution design: modules, architecture, and integration plan
Full implementation, configuration, and acceptance testing
Dataset labelling and model training on your real footage
Operator training, technical documentation, and user manual
Monthly support: tuning, continuous improvement, platform updates

Before you ask

The six questions we get every time.

Do we need to buy new cameras?

Usually not. If your cameras are ONVIF, at least 2MP at 15 FPS, and we can reach the RTSP stream, they work as they are. The assessment tells you which of your existing cameras qualify and which do not.

Does our video leave the site?

Only if you choose it to. Edge deployment keeps processing and storage on your own network. Cloud and hybrid are options, not requirements, and the decision is recorded in the design phase.

Will it work with our analog cameras?

Through an encoder, generally yes. Analog coverage is confirmed camera by camera during the assessment rather than promised in advance.

What happens when the light or the weather changes?

That is the reason models are trained on your footage and reviewed monthly. Shadows, glare, dust, and seasonal light are the conditions that break generic models, and they are what the review loop exists to absorb.

How long until we see the first alert?

Base modules can be running within hours of getting stream access. Custom models follow the training cycle, which depends on how much usable footage of the real environment exists.

What does it cost?

It depends on camera count, modules, and whether you deploy at the edge or in the cloud. We do not publish a price because we have not seen your site. The assessment produces the number, and the assessment is free.

Start here

Book the assessment. Keep the report.

One to two weeks, no cost, no commitment. You get a camera inventory, a network and RTSP readiness check, a zone and alert matrix, and a technical architecture — whether or not you go ahead with us.