Basil Alsharif

Artificial Intelligence Specialization · Cohort 2026

Hunting Leopards with Half a Brain

The Arabian leopard (Panthera pardus nimr) is critically endangered, with only a small remnant population surviving across a vast range. Conservation depends on knowing where those animals still are, but the terrain is remote and off-grid, putting real-time detection beyond the reach of cloud infrastructure. We present a split inference architecture that distributes YOLO26L computation across two Raspberry Pi devices: a Pi Zero 2W runs the backbone in the field, while a Pi 3 runs the remaining layers with post-processing. To minimise network payload, we introduce a learned convolutional bottleneck that compresses the transmitted tensor by roughly two orders of magnitude. Measured on the deployed models over a real device-to-device link, the split runs end-to-end slightly faster than the full on-device model, with minimal loss in accuracy. We demonstrate that the learned bottleneck is what makes split inference beat plain edge deployment—on payload, on accuracy, and marginally on speed—for real-time wildlife monitoring.

Read the final report (PDF)