We present CoreSight, a lightweight attendance system that combines a CNN-based 128-dimensional facial encoder with a cosine-distance matching layer to perform robust real-time student recognition. The system achieves 86.5% recognition accuracy across varied lighting, pose, and occlusion conditions on a custom classroom dataset, while sustaining ~30 FPS inference on commodity hardware. We discuss the data pipeline, evaluation protocol, and deployment considerations including privacy-preserving on-device inference.

• Designed the end-to-end pipeline: capture, encoding, matching, and persistence
• Curated and annotated a multi-condition classroom evaluation dataset
• Benchmarked encoder variants and distance metrics for accuracy / latency trade-offs
• Authored the manuscript and presented the methodology