Getting Facial-Recognition Attendance to 99%+ Accuracy in Low Light
Client: Product Build — Onest Tech LLC · Industry: HR Technology · Role: Lead Developer & AI Engineer · Duration: 4-5 months
The Challenge
Early pilots of the facial-recognition attendance system had unacceptable false-reject rates at poorly lit office entrances and warehouse floors. Employees were manually overriding the system so often it wasn't saving any admin time — undermining the entire premise of the product.
The Solution
Retrained the OpenCV/TensorFlow recognition pipeline on an internal dataset augmented with low-light and partial-occlusion samples, including masks. Added an active-lighting-check step that prompts a second capture when ambient light falls below a threshold, instead of silently failing. Layered in GPS-based liveness checks to reduce the spoofing attempts flagged during the pilot.
Results & Impact
- Recognition accuracy improved to 99.9% in target deployment conditions
- False-reject rate at low-light entrances dropped enough that manual overrides became rare
- Rolled out to 10,000+ tracked employees across client sites
- Mask detection and anti-spoofing added without a hardware upgrade