AI Retail Compliance & Planogram Auditing Platform | UK

Vision AI Planogram Auditing, Store Compliance Scoring & Regional Retail Operations — Enterprise Case Study

Project Overview

We developed a Multi-Tenant Retail Operations & Visual Merchandising Compliance Platform designed to help retail brands monitor store-level execution, verify planograms, detect out-of-stock products, and measure visual merchandising compliance at scale.

The platform combines mobile shelf photography, computer vision, AI-based image analysis, task management, escalation workflows, and executive dashboards into a centralized retail operations system.

Store managers and field teams can capture photographs of retail shelves and displays directly from mobile devices. The AI processing pipeline analyzes the images against expected merchandising configurations and identifies potential issues such as missing products, incorrect shelf placement, and visual merchandising deviations.

The platform follows a multi-tenant architecture, allowing different retail organizations, regions, stores, teams, and users to operate securely within the same SaaS infrastructure while keeping their data logically isolated.

The core operational workflow is:

Capture → Analyze → Score → Assign → Correct → Escalate → Verify → Report.

 

---

Business Challenge

Large retail organizations may operate hundreds or thousands of stores across different cities and regions.

Maintaining consistent visual merchandising across every location can be difficult.

Retail headquarters may define:

• Product placement

• Shelf layouts

• Planograms

• Promotional displays

• Product visibility requirements

• Brand standards

• Out-of-stock expectations

• Display compliance rules

However, store-level execution can differ significantly from the expected standard.

A store may have:

• Missing products

• Incorrect shelf placement

• Empty facings

• Incorrect product arrangement

• Promotional displays that are not implemented

• Products placed in the wrong position

• Poor shelf presentation

• Non-compliant merchandising

Traditionally, field teams may need to manually inspect photographs and spreadsheets to determine whether stores are following merchandising requirements.

This becomes increasingly difficult as the number of stores increases.

The client needed a platform that could automate a significant portion of this process.

---

The Solution

We developed a centralized retail operations platform that connects mobile store inspections with AI analysis and enterprise dashboards.

The architecture was designed around:

Store Manager / Field Team
          ↓
Mobile Camera
          ↓
Shelf / Display Photo
          ↓
Vision AI
          ↓
Compliance Analysis
          ↓
Compliance Score
          ↓
Task Creation
          ↓
Store Manager
          ↓
Correction
          ↓
Re-Inspection
          ↓
Regional Dashboard

This transforms retail compliance from a primarily manual process into a measurable digital workflow.

---

Multi-Tenant SaaS Architecture

A major requirement was supporting multiple retail organizations from a shared platform.

The system follows a multi-tenant architecture.

                    PLATFORM
                       │
       ┌───────────────┼───────────────┐
       ▼               ▼               ▼
   Tenant A         Tenant B        Tenant C
       │               │               │
   Regions           Regions         Regions
       │               │               │
    Stores           Stores          Stores
       │               │               │
    Users            Users           Users

Each tenant can have its own:

• Stores

• Regions

• Users

• Roles

• Planograms

• Compliance rules

• Tasks

• AI results

• Reports

• Dashboards

Tenant isolation is enforced at the application and database-access layers.

---

Retail Hierarchy

The system can represent an enterprise retail structure such as:

Organization
   ↓
Region
   ↓
District
   ↓
Store
   ↓
Department
   ↓
Shelf / Display
   ↓
Product

This hierarchy makes it possible to analyze compliance at multiple levels.

For example:

Global → Region → City → Store → Shelf → Product

---

Planogram Compliance

One of the primary features is planogram compliance analysis.

A planogram defines how products are expected to be positioned on a shelf or display.

For example:

EXPECTED PLANOGRAM

[Product A] [Product B] [Product C]
[Product D] [Product E] [Product F]
[Product G] [Product H] [Product I]

The store photograph can then be compared against the expected arrangement.

---

Shelf Photo Capture

Store managers can capture shelf photographs using the mobile application.

A typical workflow is:

Open Inspection
      ↓
Select Store
      ↓
Select Department
      ↓
Select Shelf
      ↓
Capture Photo
      ↓
Submit for AI Analysis

The camera experience is designed to make consistent photo capture easier for field teams.

---

iOS Camera Integration

For iOS, AVCapture can be used for camera capture and image acquisition.

The application can control:

• Camera preview

• Image capture

• Focus

• Exposure

• Orientation

• Image resolution

The captured image is then passed into the computer-vision pipeline.

---

CameraX Support

Where an Android implementation is required, CameraX provides the corresponding Android camera framework.

This gives the overall platform flexibility to support both:

iOS → AVCapture

Android → CameraX

while keeping the backend AI and compliance architecture consistent.

---

Vision Framework

Apple's Vision Framework was used for image analysis functionality within the Apple ecosystem.

It can support capabilities such as:

• Image classification

• Feature extraction

• Feature matching

• Object-related image analysis

• Image processing workflows

For retail auditing, these capabilities can help compare captured shelf images against expected visual configurations.

---

Visual Feature Matching

Visual feature matching can be used to compare image characteristics between:

Expected Display

and

Actual Store Display

A simplified workflow is:

Expected Reference
       ↓
Feature Representation
       ↓
Actual Shelf Image
       ↓
Feature Comparison
       ↓
Similarity / Compliance Signal

This can provide an additional signal for merchandising compliance.

---

AI Out-of-Stock Detection

One of the major AI capabilities is identifying potential out-of-stock conditions.

For example:

Expected:
[Product A] [Product B] [Product C]

Actual:
[Product A] [Empty] [Product C]

The system can flag:

Potential Out-of-Stock: Product B

This can then create a task for the store team.

---

Product Presence Detection

The AI pipeline can analyze shelf images to determine whether expected products appear to be present.

The conceptual workflow is:

Shelf Image
     ↓
Image Processing
     ↓
AI Model
     ↓
Product / Visual Detection
     ↓
Expected vs Actual
     ↓
Compliance Result

---

Visual Merchandising Compliance

Compliance is not limited to product availability.

The system can evaluate multiple dimensions such as:

• Product presence

• Product arrangement

• Shelf positioning

• Display configuration

• Promotional placement

• Visual consistency

• Planogram adherence

The exact scoring criteria can be configured for each tenant.

---

Compliance Scoring

Each inspection can receive a compliance score.

For example:

Store Compliance

████████████████░░░░
82%

Status:
Needs Attention

A score can be calculated from multiple inspection criteria.

For example:

Product Availability     90%
Planogram Accuracy       78%
Display Compliance        85%
Shelf Presentation        75%

Overall Score             82%

The exact weighting can be configured according to the organization's merchandising strategy.

---

Compliance Status

The platform can classify inspections into states such as:

• Compliant

• Mostly Compliant

• Needs Attention

• Non-Compliant

• Critical

This makes large-scale monitoring easier for regional management.

---

AI Confidence

AI predictions can include confidence information.

For example:

Product Detected
Confidence: 96%

Expected Position
Confidence: 89%

Potential OOS
Confidence: 92%

Low-confidence findings can be routed for human verification.

This creates a human-in-the-loop AI architecture rather than treating every computer-vision prediction as definitive.

---

Human Verification

The store manager or authorized reviewer can review AI findings.

The workflow is:

AI Finding
    ↓
Review
    ↓
Accept
OR
Correct
    ↓
Final Result

This is especially important for unusual shelf layouts, poor photographs, occluded products, or visually ambiguous displays.

---

Store Inspection Workflow

A complete store inspection can follow:

1. Select Store
        ↓
2. Select Department
        ↓
3. Select Planogram
        ↓
4. Capture Shelf Photo
        ↓
5. Upload / Process
        ↓
6. AI Analysis
        ↓
7. Compliance Score
        ↓
8. Review Findings
        ↓
9. Create Tasks
        ↓
10. Correct Issues
        ↓
11. Re-Capture
        ↓
12. Verify Compliance

---

Automated Task Assignment

When an issue is identified, the platform can automatically create a task.

Example:

Issue:
Product B Missing

Store:
London Store #104

Priority:
High

Assigned To:
Store Manager

Due:
Today

This converts AI findings into operational actions.

---

Task Management

Store managers can have a task list containing:

• Task

• Store

• Issue

• Priority

• Due date

• Status

• Assigned user

• Evidence

• Resolution

Typical statuses include:

Open
 ↓
In Progress
 ↓
Resolved
 ↓
Verified

---

Escalation Workflows

Not every issue is resolved immediately.

The platform therefore supports escalation workflows.

For example:

Issue Detected
      ↓
Assigned to Store Manager
      ↓
No Resolution
      ↓
Deadline Reached
      ↓
Escalate to District Manager
      ↓
Further Delay
      ↓
Regional Escalation

This allows headquarters to maintain operational accountability.

---

Task Priority

Issues can be prioritized according to business impact.

Example:

Critical

Major compliance failure or high-priority promotion issue.

High

Significant product availability or planogram problem.

Medium

Standard merchandising deviation.

Low

Minor presentation issue.

---

Regional Compliance Dashboard

The platform includes executive-level dashboards for regional management.

A dashboard can provide:

REGIONAL RETAIL COMPLIANCE

Stores Audited        248
Average Compliance    87%
Critical Issues        12
Open Tasks              38
Resolved Today          61

This allows executives to understand the operational condition of their retail network without manually reviewing every store.

---

Store-Level Dashboard

Individual stores can see their own performance.

Example:

STORE #104

Compliance Score
91%

Planogram
94%

Availability
88%

Open Issues
4

Resolved This Week
18

This provides direct feedback to store teams.

---

Regional Comparison

Management can compare different regions.

Region A     93%
Region B     88%
Region C     81%
Region D     90%

Lower-performing regions can be investigated further.

---

Store Ranking

The system can rank stores by compliance.

Example:

Top Stores

1. Store #101    98%
2. Store #203    96%
3. Store #117    95%

Needs Attention

Store #314       71%
Store #287       68%

This can help management identify both high performers and locations requiring intervention.

---

Compliance Trends

Historical data can show whether compliance is improving or declining.

Example:

January       79%
February      82%
March         85%
April         87%
May           89%

This allows executives to evaluate whether operational interventions are producing measurable improvements.

---

PostgreSQL Database Architecture

PostgreSQL was used for persistent transactional data.

Potential core entities include:

Tenants
Users
Roles
Regions
Stores
Departments
Planograms
Products
Shelf Audits
AI Findings
Compliance Scores
Tasks
Escalations
Reports

Tenant identifiers are associated with relevant records to support secure multi-tenant data separation.

---

Multi-Tenant Data Isolation

The system needs to ensure that users from one retail organization cannot access another organization's data.

A simplified data model is:

Tenant ID
   ↓
Region ID
   ↓
Store ID
   ↓
Inspection ID
   ↓
AI Findings

Every API request is authorized against the user's tenant and role.

---

Redis Cache

Redis was used to improve performance for frequently accessed or rapidly changing information.

Potential Redis use cases include:

• Dashboard caching

• Compliance summaries

• Session information

• Task counters

• Rate limiting

• Frequently accessed store data

• Temporary AI processing state

This reduces unnecessary database queries for high-frequency dashboard operations.

---

Django REST Framework Backend

The backend was developed using Python and Django REST Framework.

DRF provides the API foundation between the mobile applications, dashboards, AI services, and data layer.

Potential API domains include:

/api/auth/
/api/tenants/
/api/stores/
/api/products/
/api/planograms/
/api/audits/
/api/compliance/
/api/tasks/
/api/reports/

---

AI Processing Architecture

The AI architecture combines mobile-side image analysis with model-based processing.

A simplified flow is:

Mobile Image
     ↓
Image Preprocessing
     ↓
Vision Framework
     ↓
TensorFlow Model
     ↓
Core ML Conversion
     ↓
AI Inference
     ↓
Detected Products / Features
     ↓
Compliance Engine

This allows models to be optimized for mobile deployment where appropriate.

---

TensorFlow Model Development

TensorFlow was used to develop machine-learning models for image analysis.

Models can be trained around the visual characteristics relevant to the client's retail environment.

Potential model tasks include:

• Product recognition

• Shelf classification

• Visual compliance

• Product presence

• Display recognition

---

Core ML Model Conversion

For iOS deployment, compatible TensorFlow models can be converted into Core ML models.

The workflow is:

TensorFlow
     ↓
Trained Model
     ↓
Model Conversion
     ↓
Core ML
     ↓
iOS Application

This allows inference to occur on supported Apple devices.

---

Edge AI Processing

On-device inference can reduce dependency on continuous cloud connectivity.

The workflow becomes:

Camera
  ↓
iOS Device
  ↓
Core ML
  ↓
AI Result
  ↓
Compliance Calculation

Only required information needs to be synchronized with the backend.

---

Azure Blob Storage

Azure Blob Storage was used for large media objects such as:

• Shelf photographs

• Inspection evidence

• Reference planogram images

• Audit documentation

• Other media assets

Keeping large files outside PostgreSQL reduces database storage pressure.

---

Image Storage Architecture

The system separates structured data from media.

PostgreSQL
   │
   ├── Store Data
   ├── Audit Data
   ├── AI Results
   └── Compliance Scores

Azure Blob
   │
   ├── Shelf Photos
   ├── Evidence
   └── Reference Images

PostgreSQL stores metadata and references while Azure Blob Storage handles the large binary objects.

---

Planogram Management

Administrators can manage planograms associated with stores or product groups.

A planogram record may include:

• Planogram ID

• Store type

• Department

• Product positions

• Expected products

• Reference image

• Effective date

• Version

This allows different stores or campaigns to use different merchandising configurations.

---

Planogram Versioning

Retail layouts change frequently.

The system can therefore support planogram versions.

Planogram v1
     ↓
Planogram v2
     ↓
Planogram v3

Each inspection can reference the planogram version that was active at the time.

This improves audit accuracy.

---

Campaign Compliance

The platform can also support promotional campaigns.

For example:

Campaign:
Summer Promotion

Required:
- Promotional Stand
- Product A
- Product B
- Price Display

Store Inspection
        ↓
AI Verification
        ↓
Campaign Compliance Score

This provides a way to measure execution of marketing campaigns at store level.

---

Executive Analytics

Executives can analyze:

• Overall compliance

• Regional performance

• Store performance

• Product availability

• Planogram adherence

• Open issues

• Resolution rates

• Escalations

• Compliance trends

This turns raw shelf photographs into operational intelligence.

---

Operational KPIs

Potential KPIs include:

Compliance Rate

Percentage of inspections meeting required standards.

Out-of-Stock Rate

Percentage of expected products detected as missing.

Task Resolution Rate

Percentage of assigned issues resolved.

Average Resolution Time

Average time from issue creation to resolution.

Regional Compliance

Average compliance across a region.

Store Compliance

Individual store performance.

---

AI-to-Action Workflow

A central architectural objective was connecting AI detection directly to operational workflows.

Instead of:

AI Finds Problem

the system performs:

AI Finds Problem
      ↓
Classifies Issue
      ↓
Calculates Severity
      ↓
Creates Task
      ↓
Assigns Manager
      ↓
Tracks Resolution
      ↓
Escalates if Required
      ↓
Re-Inspects

This makes the AI system operational rather than merely analytical.

---

Development Process

Phase 1 — Retail Workflow Analysis

We mapped the client's retail hierarchy, merchandising requirements, store workflows, planogram structures, inspection procedures, and reporting requirements.

Phase 2 — Multi-Tenant Architecture

Tenant, region, store, role, and permission structures were established.

Phase 3 — Mobile Application

Swift and SwiftUI were used to create the mobile inspection experience.

Phase 4 — Camera Integration

AVCapture and CameraX support were established for consistent shelf-photo capture across supported mobile platforms.

Phase 5 — AI Model Development

TensorFlow-based computer-vision models were developed around relevant retail visual patterns.

Phase 6 — Core ML Deployment

Models were converted and optimized for supported iOS devices.

Phase 7 — Compliance Engine

AI findings were converted into planogram, availability, and merchandising compliance signals.

Phase 8 — Task Management

Automated task creation, assignment, status tracking, and escalation workflows were implemented.

Phase 9 — Executive Dashboard

Regional and store-level compliance analytics were developed.

Phase 10 — Cloud Media Storage

Azure Blob Storage was integrated for shelf photographs and other inspection evidence.

Phase 11 — Performance Optimization

PostgreSQL queries were optimized and Redis caching was introduced for frequently accessed operational data.

Phase 12 — Testing

The platform was tested across:

• Image capture

• AI detection

• Planogram comparison

• Out-of-stock detection

• Compliance scoring

• Task assignment

• Escalation

• Multi-tenant authorization

• Dashboard analytics

• Media storage

• API performance

• Database operations

• Mobile UI

---

Key Technical Challenges

Challenge 1 — Different Stores Have Different Visual Conditions

Lighting, camera angles, shelf arrangements, and product positioning can vary between stores.

The AI pipeline therefore needs to tolerate real-world visual variation.

Challenge 2 — AI Accuracy

Retail products can have similar packaging or partial occlusion.

The system therefore uses confidence scores and human verification rather than assuming every prediction is correct.

Challenge 3 — Multi-Tenant Security

A SaaS platform serving multiple retail organizations must isolate tenant data.

Tenant-aware authorization and database access controls are therefore fundamental to the architecture.

Challenge 4 — Large Numbers of Images

Shelf auditing can generate significant volumes of photographic evidence.

Azure Blob Storage was therefore used for large media files instead of storing images directly in PostgreSQL.

Challenge 5 — Turning AI Into Operations

Identifying an out-of-stock product is only useful if someone takes action.

The platform therefore connects AI findings directly to task assignment, escalation, resolution, and re-inspection.

---

Business Impact

The platform provides a digital operating layer between retail headquarters and individual stores.

It can help retail organizations:

• Automate visual merchandising audits

• Detect potential out-of-stock products

• Measure planogram compliance

• Standardize store inspections

• Assign corrective tasks automatically

• Escalate unresolved issues

• Compare regional performance

• Track store compliance trends

• Centralize photographic evidence

• Reduce manual audit work

• Turn visual data into operational intelligence

The core transformation is:

Manual shelf inspection → AI-powered visual audit → automated corrective workflow → executive retail intelligence.

---

Why Swift & SwiftUI Were Used

Swift and SwiftUI provide a native foundation for iOS field applications requiring direct access to camera and machine-learning capabilities.

The architecture is well suited to applications involving:

• Camera capture

• Vision Framework

• Core ML

• Offline workflows

• Interactive inspection interfaces

• Enterprise authentication

---

Why Vision Framework Was Used

The Vision Framework provides Apple-native image-analysis capabilities.

For this platform, it can support:

• Image classification

• Feature extraction

• Feature matching

• Visual analysis

It forms part of the mobile computer-vision layer before compliance results are produced.

---

Why TensorFlow Was Used

TensorFlow provides a flexible framework for developing custom machine-learning models.

The client-specific AI models can be trained around the products, displays, and visual conditions relevant to the retail environment.

---

Why Core ML Was Used

Core ML allows compatible models to run directly on Apple devices.

This provides:

• Low-latency inference

• Reduced cloud dependency

• Better responsiveness

• Potential offline processing

• More efficient mobile AI workflows

---

Why Django REST Framework Was Used

Django REST Framework provides a mature Python backend architecture for enterprise APIs.

It is suitable for managing:

• Authentication

• Tenants

• Stores

• Products

• Planograms

• Audits

• Tasks

• Compliance

• Reports

---

Why PostgreSQL Was Used

PostgreSQL provides reliable structured storage for the platform's transactional and relational data.

It is particularly appropriate for relationships between:

Tenants → Regions → Stores → Planograms → Audits → AI Findings → Tasks.

---

Why Redis Was Used

Redis provides fast access to frequently requested or temporary information.

For this platform, it can reduce dashboard latency and support real-time operational counters, caching, rate limiting, and temporary processing state.

---

Why Azure Blob Storage Was Used

Retail image auditing generates large amounts of photographic evidence.

Azure Blob Storage provides scalable object storage for shelf images and inspection media while PostgreSQL maintains the associated structured metadata.

---

Future Expansion

The platform can be expanded with:

• AI product recognition

• Advanced shelf-space optimization

• Price-tag verification

• Promotional compliance detection

• Competitor shelf analysis

• AI-generated merchandising recommendations

• Automated replenishment recommendations

• Predictive out-of-stock forecasting

• Retail heatmaps

• Store performance benchmarking

• Mobile offline inspections

• Voice-based task updates

• Automated regional reports

• Power BI integration

• SAP integration

• Oracle integration

• Retail IoT integration

• AI merchandising assistant

A future predictive inventory workflow could be:

Shelf Image
     ↓
AI Product Detection
     ↓
Out-of-Stock Signal
     ↓
Historical Sales Data
     ↓
Demand Analysis
     ↓
Replenishment Recommendation
     ↓
Inventory System
     ↓
Store Action

This would extend the platform from compliance monitoring into AI-powered retail operations optimization.

---

Technology Stack

iOS / Mobile

• Swift

• SwiftUI

• Vision Framework

• Core ML

• AVCapture

Android Camera Support

• CameraX

AI / Computer Vision

• TensorFlow

• Core ML Model Conversion

• Image Classification

• Feature Matching

• Visual Compliance Analysis

• Out-of-Stock Detection

Backend

• Python

• Django

• Django REST Framework

• REST APIs

Cloud Storage

• Azure Blob Storage

Database

• PostgreSQL

Performance

• Redis Cache

Core Platform Features

• Multi-Tenant SaaS

• Retail Store Management

• Planogram Management

• Shelf Photo Capture

• AI Visual Inspection

• Out-of-Stock Detection

• Merchandising Compliance Scoring

• Task Assignment

• Escalation Workflows

• Store Dashboards

• Regional Dashboards

• Executive Analytics

• Audit History

• Image Evidence Storage

---