Urgent Facial Liveness Fix

Our bespoke mobile facial-recognition login is stumbling on its liveness step. Even in normal, well-lit conditions the check often fails, leaving genuine users unable to pass authentication—and we have no fallback path, so the flow must be bullet-proof. I’ll give you access to the existing codebase, test accounts, and sample footage that reproduces the issue. Your task is to: • Pinpoint why the liveness routine intermittently refuses to validate real faces. • Patch the logic or model so legitimate users consistently pass while spoof attempts are still rejected. • Supply a concise report of the root cause, the fix, and any changes to configuration or thresholds. • Demonstrate reliability with a short test suite (recordings in good lighting across at least two major browsers). The stack is JavaScript on the front end, with the liveness inference running in a Python microservice (TensorFlow + OpenCV). If you prefer another CV library for a tighter fix, I’m open, as long as deployment remains straightforward. I’ll consider the job complete once the liveness test succeeds on 10/10 real-user tries under the same conditions that currently fail, without introducing false accepts. Fast turnaround is a priority; please be ready to dive in as soon as I share the repo. Budget: USD 10–30