Another great learning for me was to get the near real-time object detection working.
It is using a YOLOv8 model that is compiled (tflite) for the Tachyon using qai-hub. For inference, a Python script running on the Tachyon directly calls the tflite-model and processes the output. Exactly as I wanted.
This posting describes it as part of a test of the Tachyon: Particle Tachyon Review - A Qualcomm QCM6490 Edge AI and 5G cellular SBC tested with Ubuntu - CNX Software