Nikhil Deshpande

Industrial vision · prototype

Visual QA

It has never run on a shop floor, and until it does I am not claiming it works.

Local prototype · adapters parked until a factory pilot

A phone camera framing a row of moulded plastic parts on a workbench; the part in the viewfinder has a chipped edge.

A friend asked what mechanism could remove human intervention from quality checking in a plastic moulding plant. Then the same question arrived from a textile factory. Then from a garment factory. Three industries, independently, one unmet need.

What exists

A few-shot visual quality assistant for small and mid-sized manufacturers. Show it eight to sixteen photographs of a good part, and it flags parts that do not look like them and points at where. No per-product model training, no GPU, no cloud: a laptop and a phone.

Built so far: dataset preparation, tiling, frozen-backbone patch features, few-shot product profiles with nearest-neighbour and colour scoring, an evaluation harness, a capture-robustness experiment, and a local phone app. Per-industry adapters are implemented and deliberately parked until after a factory pilot.

What does not exist

A shop-floor deployment. Accuracy on real production lines. A customer. The gap between a prototype that works on a bench and a system a supervisor trusts is the whole job, and this page will say when that gap has been crossed.