Most retail AI is sold as a point solution bolted onto one device type. Edgify is the layer underneath all of them: a single control plane for the scales, scanners, self-checkouts, POS terminals, cameras and handhelds already on your shop floor.

The Edgify management console running alongside a store aisle — One control plane across every device in the estate, whoever built it.

The four layers

Layer	What it does
Application & Agent Integration	Open APIs, workflow triggers, retailer and vendor applications, and an MCP server for AI agents.
Edgify MLOps	Model lifecycle: training, validation, inference, monitoring, staged rollout and OTA updates.
Unified Real-World Data	Events, telemetry and context, under retailer-controlled governance and residency rules.
Device & Data Orchestration	Discovery, onboarding, configuration, secure communications and fleet management.

On-device training, not just inference

Machines learning together, illustrating federated model training — Each device learns from its own store. What travels between stores is the learning, never the imagery.

Running a model on the device is the easy half. The hard half is improving it once it meets the messy reality of a specific store, local produce, local packaging, local lighting, local shopper behaviour. Edgify trains on the device itself, using spare compute on hardware you already paid for.

Every customer and staff confirmation at a checkout is a label. The model learns from that label locally. What leaves the store is the learning, not the imagery.

Why that matters commercially

No annotation vendor, and no annotation budget line.
No raw shopper data crossing a network boundary, which is what makes the privacy claim defensible rather than rhetorical.
Accuracy improves after go-live instead of decaying, because the training set is the live environment.

The MCP server layer

An AI agent can already read your ERP, your ticketing system and your BI dashboards. What no agent can see is the physical store: what was picked up, weighed, scanned, mislabelled, returned to shelf or left in a trolley.

Edgify exposes that operational reality through a Model Context Protocol server, so agents query the estate as structured tools rather than scraping screens. The same permissioning, residency and governance rules apply as everywhere else on the platform. An agent sees events and state, never raw imagery.

Representative tools

list_devices: estate inventory, health and model version by site
get_events: recognition and intervention events over a time window
model_status: rollout progress, validation results and drift signals
intervention_log: what was flagged, what the shopper did next, what was recovered

For an acquirer or a platform partner, this is the part that compounds: the estate becomes queryable infrastructure rather than a closed appliance.

Hardware coverage

Edgify runs on 80%+ of existing retail hardware across 40+ OEMs. Where a device is too constrained to train locally, it still participates in inference and receives OTA model updates like the rest of the fleet.

One platform underneath every application