I have a Kotlin-based Android face-recognition app that uses the open-source FaceNet model. It performs well at first, but once the gallery grows to roughly 100–200 photos the algorithm starts matching the wrong person. The problem shows up most often when a user’s face is captured at extreme angles, so I’m convinced the root cause lies in how I generate or compare embeddings rather than in memory or UI code. What I need is a focused algorithm enhancement: • Audit the current embedding pipeline (FaceNet inference, vector normalisation, distance/threshold logic). • Adjust or replace the comparison strategy so misidentification no longer appears, even with angled faces and a larger gallery. • Keep everything in Kotlin and within the existing Android Studio / TensorFlow Lite tool-chain. Acceptance criteria 1. After your fix, the app must correctly identify all users in my existing 200-image test set, including those taken at extreme angles. 2. No new regressions in speed or APK size. 3. Clear, commented code plus a brief README explaining the changes so I can maintain them. Send changes as a pull request or patch; I’ll run the same tests that currently fail and sign off as soon as they pass.