# Store & Workforce Vision

> A privacy-aware store operations prototype for occupancy, queues, and zone events, with configurable alerts and manager review.

Turn camera events into reviewable context.

- URL: https://weaveward.com/work/retail-cctv-ai
- Stage: Prototype
- Year: 2026
- Industry: Retail and cafe operations
- Context: Internal product
- Capabilities: Computer vision, AI solutions, Business systems, Data & dashboards
- Technologies: Python, FastAPI, OpenCV, RF-DETR model adapter, ByteTrack, React, TypeScript, PostgreSQL, SQLite edge outbox, NATS

## The challenge

A store can have several cameras and still lack a useful view of its operations. Managers cannot watch every feed continuously, and hours of footage do not immediately explain a queue, an overcrowded area or a zone that needs attention.

The workflow also needs restraint. Automated activity alerts should help a manager investigate an event, with privacy controls and context before any conclusion about a person.

## What we built

Store & Workforce Vision connects edge camera analysis to a bilingual operations dashboard.

- **Events instead of endless viewing.** Configurable occupancy, queue, line and zone rules produce reviewable events. Phone-use logic includes device and exemption zones to account for approved work devices and break areas
- **An alert workflow.** Managers can search events, inspect the associated evidence and manage alerts rather than relying on a continuous feed
- **Operational reporting.** Aggregate analytics show movement, occupancy and queue patterns, with reports and optional summaries built from structured metrics
- **Camera and rule administration.** Teams configure cameras, privacy masks, zones, schedules and thresholds from the web application
- **Local resilience.** The edge component evaluates rules near the cameras and buffers delivery through a durable offline outbox

## How it works

### Privacy masks before analysis

Privacy masks are applied before analysis and storage. Stored event images and clips blur heads and faces; the system does not use face recognition. Aggregate workforce reporting is the default, with individual views restricted by policy and permission.

### Gated summaries and blurred media

Optional language-model summaries receive structured events or aggregate metrics through a gate that rejects images and binary media. The intended cloud workflow receives events, metrics and blurred event media; live viewing is separately permissioned and audited.

## Current stage

This is a substantial internal engineering prototype, exercised with a synthetic cafe simulator and local substitutes. The gallery shows rendered simulator scenes with stylised avatars, not real CCTV incidents or real people.

Real cameras, production cloud services and pilot-trained models have not been validated in the documented runs. Alert latency, real-footage accuracy and edge GPU sizing require a controlled pilot before production commitments. Managers retain responsibility for reviewing alerts and interpreting the context.

## Where this fits

Retail and hospitality operators can turn relevant camera events into a review workflow and aggregate operational insight. Explore our [computer vision](/services/computer-vision), [analytics dashboards](/services/data-analytics-dashboards) and [retail solutions](/industries/retail), or [discuss your camera workflow](/contact).