Why every flag shows a confidence score instead of a simple yes or no.

A model that just says "problem" or "no problem" asks a warehouse team to trust it blindly. We'd rather show our work.

Early on, we debated whether to simplify the model's output into a plain yes-or-no alert. It would have made for a cleaner-looking dashboard. We decided against it.

Trust Has To Be Earned, Not Assumed

We're an early-stage product working with a small number of pilot operators, and our model is still being tuned against real warehouse photos rather than a huge labeled dataset. It will get some calls wrong. A confidence score, alongside a plain-language description of what the model saw, lets a reviewer decide how much weight to give a given flag.

A 92% confidence flag on a near-empty shelf facing is a different kind of signal than a 61% flag on a pallet that's partially obscured by a forklift. Collapsing both into the same "alert" would throw away information a reviewer actually needs.

“We'd rather a reviewer occasionally overrule a low-confidence flag than trust a system that never tells them how sure it is.”

What This Means Day to Day

Every flag in the review queue carries its confidence score and a short explanation, and every reviewer decision, confirmed or dismissed, feeds back into how we tune the model. That loop is slower than shipping a black-box alert, but it's the only way we know to build something warehouse teams can actually rely on.