Interchange: change here between the ML and Hardware lines.
PCB Drishti Pro
Why
Factories check circuit boards by eye, or with rules that need every new board programmed by hand and break when the lighting changes. Every missed defect ships a bad board. Every false alarm scraps a good one. Both cost money, and nobody was showing the supervisor the cost.
How
I led a team of four. We trained a detector on photos of boards to find six kinds of fabrication defect without a reference board, and put a rupee price on every decision: release, hold for rework, or scrap. It shipped as a live service on Tata Communications cloud with one-command rollback, and we validated it on board layouts the model had never seen, so the result is generalisation rather than memory.
What came out
A national win, a live inspection service, and a decision layer that shows the cost of every call instead of just a box on a photo.
What broke
The first model failed loudly on boards that looked nothing like its training set: a photo of a bare Arduino raised 41 false alarms. Teaching it what a background looks like brought that down to almost none. The demo also keeps one bundled sample the model still misses, on purpose, because a demo that only shows the wins is a sales pitch.