Unblink builds a digital twin of your factory so vision-language models can see what is actually happening on the floor.
Most of a shift goes dark the moment it ends. A near-miss is remembered, not recorded. A line slowed down for a reason nobody wrote down. By the shift report, the texture is gone and only the totals remain.
Plants are already full of cameras. The footage is unwatched, overwritten, and reviewed only after something has gone wrong.
Modern vision models can tell you a forklift is near a person. They cannot tell you it is in *your* pedestrian zone, or one meter from the press kill-zone, or the same forklift camera 7 saw thirty seconds ago. There is no shared room they both look into.
This is the central idea: a factory becomes observable when an AI already knows the shape of the room.
We Scan Once
Engineers come to your plant. They scan the building — every aisle, fixture, machine, and camera angle — into a four-dimensional model.
The scan is durable. It does not need to be redone every shift. It is updated when the floor changes — a cell is moved, a new line is added — and is otherwise left alone.
We do not install new cameras or change your PLCs. The infrastructure that already runs your plant keeps running. The twin sits underneath everything as a calibrated, shared coordinate system: meters mean meters, *aisle 4* means aisle 4, *the press kill-zone* means a specific volume in space — labeled by a human, not guessed by a model.
What the Model Watches For
Safety is concrete. A pedestrian in a forklift path. A guard removed from a press. A spill at the entrance to a cell. PPE missing where it should not be. The model surfaces these in the moment, not in a monthly audit.
Performance has the same shape. A station idle for nine minutes between cycles. A tote stranded at the wrong buffer. A changeover that ran twelve minutes past standard.
The twin is what makes any of this trustworthy. The forklift seen by camera 3 and the forklift seen by camera 7 are the *same forklift*, because both cameras are anchored in the same world. The pedestrian zone is a labeled volume, not a guess from visual context. When the model flags something, you can open the twin and see exactly where it happened, in meters.
A request is not a command. The plant manager decides which signals become alerts, which become reports, and which stay quiet.
New Checks Are Sentences
A new safety check, in the old world, is a project. Collect footage. Label it. Train a model. Validate. Deploy. Weeks at best. Months when something goes wrong.
In this world a new check is a sentence and a polygon. Mark the volume in the twin — *the unloading dock*. Write the rule for the VLM — *flag any person standing in this volume for more than ten seconds while the door is open*. The check is live the same day.
This only works because *the unloading dock* is already a fact in the model. You did not have to teach the system what an unloading dock is. You had to tell it where yours is, once, when we scanned.
Performance follows the same shape. A plant manager who wants to track *time-to-changeover at line 4* does not commission a study. They name the line in the twin, describe what a changeover starts and ends with, and the measurement runs.
Factory observability gets built once, in person, and extended in language. The system understands your building because someone walked it.
A factory you can see.