AI Retail Store Safety Monitor

Real-time AI hazard detection for retail stores. Turn any Android phone into a live safety monitor — zero hardware cost, deploy in 5 minutes.

Project 5 of 5 in a portfolio of real-world AI Android apps. View full portfolio →

---

Screenshots

       

---

The Problem

• $50B+ annual retail safety losses from preventable in-store incidents
• $300,000 average cost per slip-and-fall incident — legal fees, medical bills, insurance hikes
• OSHA violations up to $156,259 per citation for willful safety violations
• Manual safety walks miss 95% of hazards — a manager can't watch every aisle at once
• No affordable solution for small/medium retailers — enterprise systems cost $50,000+ in hardware

---

The Solution

Turn any Android phone into a real-time AI safety monitor.

A phone mounted at the end of an aisle runs ML Kit Object Detection on every camera frame. Hazards are detected, classified by OSHA severity, and escalated to the manager via push notification — all without leaving the device. The Gemini 2.5 Flash weekly report turns raw incident data into actionable safety insights every Monday morning.

Zero hardware cost. Deploy in 5 minutes.

---

Features

• Real-time ML Kit Object Detection on the live camera feed, every frame, on a background thread
• Colour-coded bounding boxes — red (CRITICAL), amber (HIGH), yellow (MEDIUM), green (LOW)
• Instant push notifications for every newly detected hazard, severity-labelled
• Smart escalation engine — WorkManager re-notifies managers every 15–120 min based on severity
• Weekly AI safety report via Gemini 2.5 Flash — trends, top risks, actionable recommendations
• Compliance score dashboard — severity-weighted 0–100 score, resets every Monday
• Full incident log with hazard type, severity, timestamp, and location description audit trail
• Works offline — ML Kit runs entirely on-device, no network required for detection
• Clean Architecture + MVVM + Hilt + UDF — every layer tested independently

---

How It Works

flowchart TD
    A[Manager mounts phone, opens app] --> B

    subgraph Camera ["CameraX + Background Thread"]
        B[PreviewView + ImageAnalysis bind to lifecycle]
        B --> C[ML Kit Object Detection — every frame\ncameraExecutor single thread, never main thread]
    end

    C --> D{Confidence ≥ 0.7?}
    D -->|No| C
    D -->|Yes| E

    subgraph Detection ["Detection Pipeline"]
        E[HazardDetector maps ML Kit label → HazardType\nFood→SPILL, Home good→FALLEN_ITEM\nFashion good→TRIP_HAZARD, Place→BLOCKED_EXIT]
        E --> F[60-second per-type cooldown\nprevents Room DB flooding at 30fps]
        F --> G[DetectHazardUseCase\nmaps HazardType → Severity]
        G --> H[LogHazardUseCase\nRoom insert + push notification]
    end

    H --> I[HazardOverlay draws bounding boxes\nnormalized 0.0–1.0 coords × canvas size]
    H --> J[MonitorUiState.HazardDetected\nbottom sheet expands, 5s auto-dismiss]

    subgraph WorkManager ["WorkManager Escalation Engine — every 15 min"]
        K[HazardEscalationWorker\ngetAllUnresolvedSnapshot]
        K --> L{age ≥ escalationIntervalMs\nAND cooldown elapsed?}
        L -->|Yes| M[sendEscalationNotification\nupdateLastEscalatedAt]
        L -->|No| N[Skip — idempotency guard]
    end

    subgraph Report ["Weekly AI Report — every Monday"]
        O[WeeklyReportWorker triggers\nGenerateSafetyReportUseCase]
        O --> P[Compliance score calculation\nbase - criticalPenalty - highPenalty]
        P --> Q[Gemini 2.5 Flash\ntrend analysis + recommendations]
        Q --> R[SafetyReport persisted to Room\nReportScreen renders results]
    end

    H --> K
    H --> O

Loading

Iron rule: ML Kit inference runs exclusively on cameraExecutor. StateFlow.update{} is called directly from that thread — zero coroutine allocations per frame.

---

OSHA Severity Mapping

Severity	Example Hazard	OSHA Classification	Max Fine
🔴 CRITICAL	Blocked fire exit	Willful/Repeat violation	$156,259 per citation
🟠 HIGH	Wet floor, no signage	Serious violation	$15,625 per citation
🟡 MEDIUM	Trip hazard, fallen item	Other-than-serious	$15,625 per citation
🟢 LOW	Minor clutter, unknown object	De minimis notice	No mandatory fine

Escalation interval: CRITICAL every 15 min → HIGH every 30 min → MEDIUM every 60 min → LOW every 120 min.

---

Architecture

Clean Architecture

flowchart TB
    subgraph UI ["UI Layer (Jetpack Compose)"]
        direction LR
        MonitorScreen --> MonitorViewModel
        DashboardScreen --> DashboardViewModel
        IncidentsScreen --> IncidentsViewModel
        ReportScreen --> ReportViewModel
        SettingsScreen
    end

    subgraph Domain ["Domain Layer (Pure Kotlin — zero Android deps)"]
        direction LR
        DetectHazardUseCase
        LogHazardUseCase
        ResolveHazardUseCase
        GetHazardHistoryUseCase
        GenerateSafetyReportUseCase
        HazardRepository
        ReportRepository
    end

    subgraph Data ["Data Layer"]
        direction LR
        HazardDetector
        CameraManager
        HazardRepositoryImpl
        ReportRepositoryImpl
        HazardNotificationManager
        HazardEscalationWorker
        WeeklyReportWorker
        GeminiReportApiImpl
        SafetyDatabase
    end

    UI --> Domain
    Domain --> Data

Loading

Unidirectional Data Flow (UDF)

sequenceDiagram
    participant UI as MonitorScreen
    participant VM as MonitorViewModel
    participant Cam as CameraManager
    participant Det as HazardDetector
    participant UC as LogHazardUseCase
    participant DB as Room DB
    participant NM as HazardNotificationManager

    UI->>VM: startCamera(lifecycleOwner, previewView)
    VM->>Cam: startCamera(executor, onHazardsDetected)
    Cam->>Det: bindToImageAnalysis(imageAnalysis)
    Note over Det: Every frame on cameraExecutor thread
    Det-->>VM: onHazardsDetected(results) — StateFlow.update{}
    VM-->>UI: MonitorUiState.HazardDetected
    VM->>UC: execute(hazardType, severity)
    UC->>DB: insertHazard(hazard)
    UC->>NM: sendHazardDetectedNotification(hazard)
    UC-->>VM: Hazard
    VM-->>UI: bottom sheet expands, 5s timer starts

Loading

WorkManager Escalation Flow

sequenceDiagram
    participant WM as WorkManager
    participant W as HazardEscalationWorker
    participant R as HazardRepository
    participant NM as HazardNotificationManager

    Note over WM: Fires every 15 minutes (platform minimum)
    WM->>W: doWork()
    W->>R: getAllUnresolvedSnapshot()
    R-->>W: List<Hazard>
    loop For each unresolved hazard
        W->>W: ageMs = now - hazard.timestamp
        W->>W: cooldownElapsed = now - lastEscalatedAt ≥ interval
        alt Both conditions met
            W->>NM: sendEscalationNotification(hazard, ageMinutes)
            W->>R: updateLastEscalatedAt(hazard.id, now)
        else Cooldown not elapsed
            Note over W: Skip — idempotency guard prevents spam
        end
    end
    W-->>WM: Result.success()

Loading

---

Key Engineering Decisions

#	Decision	What	Why
1	StateFlow.update{} direct from cameraExecutor	ML Kit callbacks mutate state without launch {}	At 30fps, launching a coroutine per frame generates ~1,800 allocations/min and risks GC-induced dropped frames. StateFlow is thread-safe — no launch needed.
2	lastEscalatedAt in HazardEntity	Epoch millis of last escalation persisted in Room	Without this, a CRITICAL hazard open for 2 hours would generate 8 duplicate notifications. Per-hazard cooldown is the idempotency contract.
3	Weekly compliance score scope	Score resets every Monday 00:00	Retail operations run on a Mon–Sun cycle. A rolling 7-day window would blur weekly performance. Monday reset matches how store managers think about their numbers.
4	WorkManager = escalation only, not camera	Workers never access the camera hardware	Android 12+ (API 31+) blocks while-in-use camera access in background workers. WorkManager is the escalation and reporting engine only — CameraX requires a foreground lifecycle owner.
5	ML Kit confidence threshold 0.7	Hard-coded in HazardDetector, applied after inference	Below 0.7: too many false positives flood the incident log. Above 0.7: real hazards get missed. 0.7 is the industry-standard balance for real-time object detection on consumer hardware.

---

Tech Stack

Component	Technology
Language	Kotlin 2.2.10
UI	Jetpack Compose + Material 3
Architecture	Clean Architecture + MVVM + Hilt 2.59.1 + UDF
Camera	CameraX 1.3.4 (Preview + ImageAnalysis)
Object Detection	ML Kit Object Detection 17.0.2 (on-device)
AI Reports	Gemini 2.5 Flash via Retrofit
Database	Room 2.7.1
Background Jobs	WorkManager 2.9.0
Dependency Injection	Hilt (Dagger)
Networking	Retrofit 2.11.0 + OkHttp 4.12.0
Permissions	Accompanist Permissions 0.37.0
Image Loading	Coil 2.7.0
Testing	JUnit4 + Mockito-Kotlin 5.4.0 + Turbine 1.2.0
CI/CD	GitHub Actions

---

Setup

Prerequisites

• Android Studio Ladybug or later
• Android SDK 36
• Physical Android device (API 26+) — camera features require real hardware
• Gemini API key (get one free at aistudio.google.com)

Clone

git clone https://github.com/lakshmanreddymv/RetailSafetyMonitor.git
cd RetailSafetyMonitor

Configure API Key

Add to local.properties (this file is gitignored — never commit it):

gemini.api.key=YOUR_GEMINI_API_KEY

Build and Run

./gradlew assembleDebug        # build debug APK
./gradlew installDebug         # install on connected device
./gradlew test                 # run all 55+ unit tests

Note: The Gemini API key is only required for the weekly report feature. All ML Kit object detection, hazard logging, notifications, and escalation work fully offline with no API key.

---

How to Test

Test	Steps	Expected Result
1. Camera launch	Open app → grant camera permission	Live preview appears, "● LIVE" indicator, ML loading spinner then disappears
2. Hazard detection	Point camera at a wet floor sign, spill, or cluttered aisle	Colour-coded bounding box appears, bottom sheet expands with hazard card
3. Push notification	Let hazard stay on screen → check notification tray	Notification fires: "HIGH: Wet Floor Detected"
4. Compliance score	Open Dashboard tab	Gauge shows 0–100 score, stat chips show detected/resolved counts
5. Weekly report	Dashboard → "Generate Weekly AI Report"	Gemini analyses incidents, report appears in Report tab
6. Offline mode	Enable airplane mode → open Monitor tab	ML Kit detections continue — only Gemini report is blocked
7. Unit tests	./gradlew test	55+ tests pass, 0 failures

---

Unit Tests

Test Class	Tests	What It Verifies
HazardDetectorTest	14	ML Kit label → HazardType mapping, confidence threshold filtering, OVERCROWDING aggregation (≥3 persons), position heuristics
DetectHazardUseCaseTest	8	HazardType → Severity mapping, all 8 types covered, CRITICAL/HIGH/MEDIUM/LOW tiers
GenerateSafetyReportUseCaseTest	10	Compliance score formula, Gemini fallback on failure, zero-hazard edge case returns 100
ComplianceScoreTest	8	Severity-weighted penalty calculation, critical cap at 40, high cap at 20, floor at 0
DashboardViewModelTest	7	Weekly scope query, score recomputation, report generation state transitions
MonitorViewModelTest	8	Idle → Monitoring → HazardDetected → 5s auto-dismiss, pause/resume, error recovery
HazardRepositoryImplTest	6	Room insert/query, entity ↔ domain mapping, lastEscalatedAt persistence
HazardEscalationWorkerTest	4	Idempotency guard (age threshold AND cooldown both required), Result.retry on DB exception
Total	65	0 failures

./gradlew test

Test reports: app/build/reports/tests/testDebugUnitTest/index.html

---

Bugs Fixed During Build

#	Bug	Root Cause	Fix
1	ksp plugin version 2.2.10-1.0.31 not found	KSP2 version format changed — {kotlin}-{ksp2_release}	Correct version is 2.2.10-2.0.2
2	libs.mlkit.object.detection parse error	object is a Kotlin reserved keyword in TOML accessor path	Renamed alias to mlkit-detection → libs.mlkit.detection
3	Failed to resolve camera-mlkit-vision:1.3.4	androidx.camera:camera-mlkit-vision does not exist at CameraX 1.3.4	Removed it entirely; rewrote HazardDetector to use InputImage.fromMediaImage() directly
4	AndroidBaseExtension not found	Hilt 2.56 uses AGP APIs removed in AGP 9.x	Updated to Hilt 2.59.1
5	plugin kotlin.android already on classpath	kotlin.android declared in both root build.gradle.kts and app/build.gradle.kts	Removed kotlin.android from app module — kotlin.compose includes it; added apply false declarations to root
6	Unresolved reference 'surfaceProvider'	CameraX 1.3.4 uses setSurfaceProvider() method, not a property setter	Changed it.surfaceProvider = … to it.setSurfaceProvider(…)
7	Unresolved reference 'setClassificationConfidenceThreshold'	Method does not exist in ML Kit Object Detection API 17.0.2	Removed the call — confidence filtering is applied manually in processFrame()
8	OkHttpClient could not be resolved (KSP2 error)	Import was com.squareup.okhttp3.OkHttpClient — that is the Maven group ID, not the Java package	Changed to okhttp3.OkHttpClient (the actual package name)

---

Real-World Use Cases

Scenario	Detection	Action
Spilled coffee, no sign	Wet floor visible on camera	CRITICAL → immediate notification to manager, $156K OSHA fine prevented
Stock cart blocking fire door	"Place" label near screen edge → BLOCKED_EXIT	CRITICAL → escalates every 15 min until resolved
Dropped produce in produce aisle	"Food" label on floor → UNATTENDED_SPILL	MEDIUM → manager notified, escalates after 60 min
Monday morning staff meeting	Weekly report generated overnight	Gemini: "Tuesday afternoons show 3× higher hazard rate — schedule extra walk at 2pm"
OSHA surprise inspection	Inspector requests safety records	6-month incident log with timestamps, types, and resolution status exported from Room
Power outage / no WiFi	Offline store, internet down	ML Kit runs entirely on-device — detection, logging, and local notifications continue
300-location retail chain	Area manager oversight	Per-store compliance scores surfaced in weekly reports for regional benchmarking

---

Market Context

	RetailSafetyMonitor	IntelliSee / Visionify / Protex AI
Hardware required	None — any Android phone	Dedicated IP cameras ($2,000–$10,000 per store)
Setup time	5 minutes	2–4 weeks installation
Cost	Zero hardware + Gemini API (~$5/month)	$50,000+ enterprise contract
Target	SMB retailer (1–20 locations)	Enterprise retail chain (100+ locations)
Offline capable	Yes — ML Kit on-device	No — cloud-dependent

Target market: 300,000+ SMB retailers in the US with no affordable safety monitoring solution.

---

Roadmap

• Multi-store dashboard for area managers — aggregate compliance scores across locations
• On-device Gemini model — fully offline weekly reports, zero API cost
• PDF audit trail export — OSHA-ready incident report for compliance inspections
• Wearable integration — CRITICAL hazard alerts pushed to manager's smartwatch
• Custom hazard training per store type — grocery vs. hardware vs. pharmacy profiles

---

Portfolio

Project	Description	Tech
MySampleApplication-AI ✅	AI Natural Language Search	Gemini API + Clean Architecture
FakeProductDetector ✅	Counterfeit detection ($500B problem)	Gemini Vision + CameraX
EnterpriseDocumentRedactor ✅	GDPR/HIPAA PII redaction, fully offline	ML Kit OCR + on-device AI
MedicalSymptomPreScreener ✅	Safety-critical AI triage	Gemini + Three-layer safety
RetailSafetyMonitor ✅	Real-time AI hazard detection	ML Kit + CameraX + Gemini

---

License

MIT License — see LICENSE

---

Author

Lakshmana Reddy — Android Tech Lead, 12 years experience Building AI-powered Android apps | Pleasanton, California

GitHub | lakshmanreddymv@gmail.com