Computer vision for manufacturing. Without sending a single frame to the cloud.
Counting runs in production on an edge device on your plant floor, on the cameras you already have. The image is processed and discarded; what gets stored is the event, not the footage. Defect detection is trained on images from your line and PPE checks are in development.
Proven in production
The count arrived a week late. Exactly the cycle where laying posture changes.
Grupo Proa, an egg producer in Chile, counted by hand. The number reached management when it was no longer useful for any decision, and it carried a wide deviation. A camera over the line now counts every unit and the figure is available within the shift.
- Counting deviation, before
- 5–15%
- Accuracy vs. manual count
- 99%
- Data latency, before
- 7 days
- Latency now
- same shift
What it actually does on a line
Four applications on the same edge device. Counting is in production; the others depend on your line.
Automated counting
Units, cases, trays. Replaces manual tallies and clipboard counts that arrive late and carry error.
Defect detection
Surface flaws, misalignment, wrong fill level. Trained on labeled images from your own line: without that dataset there is no reliable detection.
PPE and safety (in development)
Helmet, goggles, restricted-zone access. In development; it will record the event, not the operator's face.
Label and code verification
Reads what actually printed and compares it against what should have printed, before the pallet ships.
Why the inference belongs on the plant floor
Every vendor in this space processes in the cloud. That choice has four consequences, and none of them are about ideology.
Latency
A reject decision at line speed cannot wait for a round trip to a datacenter.
Bandwidth
Streaming continuous video off-site costs more than the inference itself, every month.
Personal data
A camera that captures operators handles biometric data. If the frame never leaves, there is no transfer to justify.
Link outages
The plant keeps running when the connection drops. So does the inspection.
None of this makes the cloud wrong. It makes it the wrong place for this particular decision — the one that has to happen in milliseconds, next to the machine, on footage of your people and your product.
How on-premise processing worksReal demos, unedited
Recorded from iGromi Vision deployments and trials: counting and detection running on real footage from plants, ports, pens and highways.
What connects
Your plant's three pillars, in one place.
The security cameras already installed
Counts that reach production and dispatch
A person validates what the camera reports
- The AI reads, proposes and cites the source
- A person decides; no machine moves on its own
- The Edge reads the PLC read-only
How it works
1 · What your plant already has
- The line’s security cameras
- The PLC signal, if there is one
- Shifts and product in process
2 · The iGromi node on site
- Detects and counts every unit that passes
- Cross-checks the count with the machine signal
- Processes video on site, without uploading it
3 · What your team gets
- Units per hour, shift and line
- An alert when the pace drops
- The count, exportable to your MES or sheet
Which plan
This is part of Enterprise.
Enterprise adds your assets with an Edge inside your plant: it reads PLCs, sensors and cameras read-only and data is processed there.
iGromi's cloud, separated by company
Your own isolated cloud
Inside your plant, with Edge or on-premise
Every plan starts with the free app: your team logs from the first shift, and systems and machines join when you want.
FAQ
How is this rolled out in a plant that is already running?
The same five questions come up in every meeting, whichever solution is on the table. The answers depend on how the system works, not on the module.
Do I have to replace my PLC or my SCADA?
No. The system connects to what is already installed and reads it read-only; it does not replace the control layer and does not write to it. The AI proposes and a person decides. If the PLC does not expose its parameters, the iGromi Bridge discovers them by trying combinations until it gets a response, and if the machine speaks no protocol at all it is read by a sensor on the electrical signal.
Does it work if the internet goes down?
Yes. Processing happens at the Edge, inside the plant, so capture, rules and inference keep running with the link down. What stops is remote access and outbound synchronisation, which resumes on its own when the connection returns.
Does my data leave the plant?
Not by default. Video and telemetry are processed on the customer's Edge device. What leaves, and where it goes, is defined when the system is configured. That is what makes it possible to operate under Chile's Ley 21.719 and what makes this viable in mining and in food.
How long until the first data point?
It depends on how many machines there are and what they expose. A line with equipment that speaks Modbus is read in days; a line of silent machines needs sensors installed and panel shutdowns coordinated, and that is set by the plant's maintenance calendar, not by us. The usual sequence is assessment, pilot, expansion.
What happens to what we already measure by hand?
It runs alongside during commissioning. For the first weeks the manual record keeps going in parallel as a validation counterpart: it is how you find out whether the automatic figure beats the handwritten one, and the habit is dropped on its own once that is demonstrated.
Start with one camera on one line
We connect an existing RTSP or ONVIF camera and show you the count from a real shift. No new hardware to begin.
- 1Request early access, free
- 2Pick your trade: maintenance, production or safety
- 3Add systems and machines when you want
No credit card. Your data stays yours.