Tarush Nagpal built a local face recognition system over a weekend using an ESP32-CAM costing just ₹800. The setup streams video to a Python server that identifies individuals without cloud APIs or subscriptions. It’s a privacy-friendly presence detector, not surveillance software, designed for personal experimentation rather than production environments.

The system architecture uses MJPEG streaming from the ESP32, which the Python server consumes by scanning for JPEG markers. The face_recognition library (built on dlib) handles detection and encoding. Instead of manual labeling, the software automatically creates new person IDs for unknown faces, storing multiple encodings per person to improve accuracy.

To make the data useful, detections are bucketed into 30-second intervals, smoothing out the graphs in the Flask and Chart.js dashboard. The final result processes roughly 10 frames per second entirely on the local network. Nagpal noted that 5V power is critical for the ESP32-CAM, as 3.3V causes random reboots.