The problem
A phone that is unlocked is trusted, whoever is holding it. Checking behavior in the cloud is slow and sends private signals off the device.
What OD-MAS does
OD-MAS learns its owner’s behavioral patterns on the phone and flags behavior that does not match. Everything runs on the device, and each decision takes under 200 ms.
How it works
Detection runs in tiers. Tier 0 is a fast statistical check based on Mahalanobis distance. When it is unsure, Tier 1 asks a neural ensemble. Confidence-aware hysteresis keeps the system from flipping on a single odd reading. The behavioral models run through Chaquopy, which brings Python into the Android app.
Result
OD-MAS won the Samsung EnnovateX AI Challenge in 2025. I was the lead engineer.