Journal
The last mile is the system
Physical AI fails at the camera, the radio, and the warehouse with no signal — not in the model card.
A face-recognition slide is easy. A yard at 2 a.m. with a dirty lens, a flaky BLE badge, and a tablet that has not seen a tower in four hours is the actual product.
We have owned that last mile: ANPR and face recognition on the edge, Node ingestion, React Native and Electron field clients, a custom BLE SDK, live video, geotagging, service-worker offline. The model is a line item. The system is the cameras, the queues, the sync, and the operator who will not wait for a spinner.
What breaks first
Streams. RTSP and WebRTC do not care about your sprint. Offline-first is not a flag; it is a conflict log and a story about which write wins when the truck returns. Hardware SDKs lie about reconnect. If you have not held the device, you have not designed the API.
How we approach it
Start from the field action, not the dashboard. What must be true for a gate to open, a plate to be fined, a face to be attested. Then walk backward to the cloud. Anything that only exists in the happy-path mock is not in the system.
Physical AI is unfashionable because it is slow to demo and expensive to fake. That is why it is still a business. The last mile is the system. We build it as one.