I realized that building a robot prototype in just six weeks can still end in failure months later if the model never sees real‑world variability.
Build time: 6 weeks from concept to prototype.
Post‑deployment lag: 6 months later the system still isn’t working.
Training environment: the robot was trained on clean data from a clean lab.
- Real‑world conditions encountered: • Broken latches • Inconsistent lighting • Objects in unexpected orientations
Model exposure gap: the model never saw any of it.
I observed a micro‑delay on a hydraulic‑fittings line near Ludhiana: the seating step for the 11ᵗʰ piece took ½ second longer than usual, invisible to a supervisor who only scans the floor 4 seconds every hour.
The deviation was exactly ½ second longer on the seating step for the 11ᵗʰ piece.
That piece failed a pressure test 2 stations later.
Supervisor visibility amounts to 4 seconds every hour.
India has roughly 6.3 crore MSMEs.
I decided to build a visual‑AI agent that continuously watches the line, flags ≥ ½ second anomalies, and links them to downstream quality failures, without ranking workers.