T-14 / 11
T-22 / 11
Click any wagon to open its record.
Your cameras already see the yard, the gate and the line. Today that footage is watched by nobody and the real record lives in a spreadsheet someone retypes by hand. We turn the cameras and sensors you already own into a live operational picture — with the uncertain cases handed to a person instead of guessed.
Click any wagon to open its record.
Cockpit mock-up for dispatcher workflows. The shipped prototype draws the same picture from live camera feeds.
Anywhere something physical moves through your site and the only record of it is a person with a clipboard and a spreadsheet.
Every wagon on every track with its number, weight, cargo and dwell time. Alerts when a track goes over capacity or a wagon has stood for four hours. The shunting locomotive stops being counted twice.
Plate recognition at the gate, time in and time out, which bay, how long. The difference between "the truck was here about eleven" and a timestamped record with a photo attached.
The weighbridge reading tied automatically to the vehicle or wagon that was actually on it, with the frame that proves it. Re-weighing because the paperwork got out of order stops.
What is physically standing in the yard tonight, against what the system believes is standing there. The gap between the two is usually where the money is.
How long loading actually takes, per crew and per shift, measured rather than estimated. Stoppages detected and timed instead of reconstructed from memory the next morning.
A person in a zone that should be empty during a movement, a barrier left open, a vehicle where it should not be. Events and zones, not scoring individual workers — that distinction matters legally and it matters to your staff.
Every one of these starts the same way: an hour of your footage and a conversation about what the spreadsheet currently costs you.
Book a callSix stages, and the last one is a person. Renting somebody's recognition service at fifty dollars per camera per month still leaves your staff retyping the ones it missed — the part worth engineering is what happens when the model is not sure.
Camera projects have a reputation for being paid up front and never quite working. So each stage ends in a live demonstration on your site, and you pay for the stage after you have seen it.
The full move off Excel first, before any AI: the yard map, the journal, roles and permissions, history and audit log. Plus hardware sizing and the first camera footage collected.
Data labelled and the recogniser fine-tuned on your cameras and your wagons. It proposes, your operator corrects. Accuracy is measured on your footage, not on a benchmark.
The system runs on its own and the operator only confirms exceptions. The numbers that go into your contract are the ones measured on your data in stage two.
Weighbridge, wheel-pair sensors — including analogue gauges read through a camera — automatic reports and acceptance workflows.
A prototype built on recorded footage from an industrial site. Here is what that does and does not mean.
The pipeline has been run end to end on real footage from a working coal station. It has not yet been running continuously in a control room for a year. Anyone claiming otherwise about a system like this should be asked for the log files.
The single biggest improvement in our test came from where a camera pointed, not from the model. Sometimes the honest recommendation after looking at your site is to move two cameras and re-test before spending anything on AI.
We build zone and event detection — a movement, a barrier, an empty area that should stay empty. We do not build individual productivity scoring or face-based tracking of your staff. That is a position, not a technical limit, and it keeps the system on the right side of both the GDPR and your workforce.
Usually not for the first test — we work with a recording from what you already have. After looking at the footage we will tell you honestly whether the angles are workable, whether moving one camera would fix it, or whether a new one is genuinely needed. That conversation happens before you buy anything.
No. It runs on a server on your site. Video is the most sensitive material most industrial companies have, and sending it to a cloud API is the kind of thing that ends a project at the security review. For the initial test we agree in writing what we receive and when we delete it.
Per-camera monthly rent scales with your site while the accuracy does not improve, and your staff still retype whatever it missed. An on-premise system you own is a one-off build that gets better on your data every time an operator corrects it. If your current service works and the rent is small, keep it — we will say so.
It says so. The wagon appears in the review queue with the frame attached and the operator types what they see. There is no configuration in which the system invents a plausible number, because a wrong record is more expensive than a missing one.
The pipeline is the same: find the event, read the marking, validate it against what a valid marking can look like, vote across frames, confirm across cameras, escalate the rest. Plates and container codes have their own check rules, which makes them if anything easier than a rusted wagon stencil.
We run it and come back with what was found, what was read, what was missed and why. No commitment, and if the answer is that your cameras are in the wrong place, that is what you will hear.
Book a call with Artem