Recipes

The starting point — load a model and get bounding boxes back.

When the answer you need is "what is this" rather than "where is it".

Compare images numerically — find duplicates, build a similarity index, train your own classifier.

Where to put your key, and how to confirm the GPU is actually engaged.

A wrong key, an expired license, and a corrupt file each need different handling.

Process video of any length without running out of memory.

Watch every camera from a single inference session — better throughput than one process per stream.

Persistent IDs across frames — the foundation for counting, dwell time, and behaviors.

Expose LYNX as a web service so any language can call it.

Fire actions when something specific happens — a class detected, a zone entered, a line crossed.

Bounding boxes give you shape; depth gives you "how far away is that thing".

See exactly what input format and defaults a model expects, without reading anyone's prose.

Exact CLI behavior from inside Python — useful for reproducing CI runs in a notebook.

When the model returns nothing useful, the cause is almost always preprocessing — find it in one call.

Send us a frame the model got wrong — the next training cycle benefits from real-world failures.