Overview

Architecture

Each SCO gives the AI Box two inputs: the camera view and its transaction events. The AI Box sends one output back to the SCO: an alert.

Angible AI Box

Camera video stream

Transaction events

Real-time alerts

One camera per self-checkout

Camera 1SCO 1

Camera 2SCO 2

Camera 3SCO 3

Input — into the AI Box

Camera view · Camera → AI Box (RTSP)

The AI Box pulls one RTSP stream per SCO.

Transaction events · SCO → AI Box (REST)

The SCO pushes transaction logs as they happen.

Output — back to the SCO

Real-time alerts · AI Box → SCO (MQTT)

The AI Box compares the two inputs. When they don’t match, it publishes an alert back to that specific SCO.

Integration stages

The SCO integration runs through the stages below. They are largely sequential, except that 3a and 3b run in parallel — the alert-handling UI (3b) is built while Angible fine-tunes the model (3a).

Each build stage brings up one channel from the architecture above: stage 1 the REST transaction events, stage 2 the RTSP camera stream, and stage 3b the MQTT alerts. UAT and go-live run all three together.

#	Stage	Channel	Goal
1	Transaction event integration	REST	Build the SCO → AI Box transaction-event integration in a lab/test environment.
2	On-site data collection	RTSP	Deploy in the real store and collect real transaction events, aligned with the live camera video.
3a	AI model fine-tuning	—	Fine-tune Angible's general-purpose AI model on the store's real items and camera setup.
3b	Alert resolution UI development	MQTT	Build the SCO screens that respond to the AI Box's alerts (customer and staff resolution flows).
4	UAT testing	All three	Validate the full loop with AI end-to-end.
5	Go live	All three	The store goes into production; the integration runs live for real shoppers.