An AI-powered iOS expense tracker that extracts data from receipts using on-device OCR and machine learning. Scan a receipt, get structured expense data — no cloud processing, 100% private.
PrivExpensIA is the subject of an IEEE-format research paper (September 2025) documenting the multi-agent AI framework that built it:
"PrivExpensIA: Observational Study of a Multi-Agent Orchestration Framework Based on Local Claude CLI with Email-mediated Communication"
The paper details how 4 autonomous AI agents (Nestor, Tintin, Dupont1, Dupont2) collaboratively developed this application through sprint-based coordination, email-mediated communication, and real-time observability — two months before multi-agent frameworks became mainstream. See the full paper at docs/PrivExpensIA_Moulinsart_IEEE_v3.pdf.
PrivExpensIA combines Apple's Vision Framework with a quantized Qwen2.5-0.5B language model to extract merchant names, amounts, dates, categories, and tax information from receipt photos — all running entirely on your iPhone. No data ever leaves your device.
- OCR Engine — Vision Framework with image preprocessing, orientation detection, and result caching
- AI Extraction — Qwen2.5-0.5B (4-bit quantized, ~300MB) for intelligent field parsing
- Multi-page Scan — Process multiple receipts in a single scanning session
- < 2 seconds OCR processing, < 300ms AI inference
- Smart Categorization — AI-powered category detection (food, transport, utilities, etc.)
- Budget Tracking — Set monthly budgets and monitor spending
- Statistics Dashboard — Visual charts and spending insights
- Reports — Organize expenses into folders and generate summaries
- Invoice & Admin Documents — Separate storage for receipts vs administrative documents
- AI Classification — Automatic invoice vs. admin document detection
- AI Summaries — Auto-generated document summaries via RAG
- iCloud Sync — Cross-device document synchronization
- Built-in conversational AI for expense-related questions
- Multi-provider support for flexibility
Full support for 12 language variants:
| Language | Variants |
|---|---|
| French | FR, FR-CH (Swiss) |
| German | DE, DE-CH (Swiss) |
| Italian | IT, IT-CH (Swiss) |
| English | EN |
| Spanish | ES |
| Japanese | JA |
| Korean | KO |
| Slovak | SK |
- Dark Mode — Full support with liquid glass UI theme
- Accessibility — VoiceOver and Dynamic Type support
- Glass Effects — Modern frosted glass design language
- iOS 17.0+
- Xcode 15.0+
- Swift 5.0+
- iPhone or iPad with camera
# Clone the repository
git clone https://github.com/yourusername/privexpensia.git
cd privexpensia/PrivExpensIA\ APP
# Generate Xcode project
xcodegen generate
# Open in Xcode
open PrivExpensIA.xcodeproj
# Build and run (Cmd+R)
# Select your device or simulator as the targetNote: The Qwen2.5 model (~300MB) is included in the project. First build may take longer as SPM resolves dependencies.
PrivExpensIA APP/
├── PrivExpensIA/
│ ├── App/ # Entry point & lifecycle
│ ├── Views/ # SwiftUI views
│ │ ├── ScannerGlassView.swift # Camera & OCR scanner
│ │ ├── ExpenseListGlassView.swift # Expense list
│ │ ├── StatisticsGlassView.swift # Analytics dashboard
│ │ ├── DocumentArchiveView.swift # Document storage
│ │ └── ChatAssistantView.swift # AI chat
│ ├── Services/
│ │ ├── OCRService.swift # Vision Framework OCR
│ │ ├── QwenModelManager.swift # AI model (lazy loading)
│ │ ├── AIExtractionService.swift # Receipt data extraction
│ │ ├── CoreDataManager.swift # Persistence layer
│ │ ├── DocumentSyncService.swift # iCloud sync
│ │ └── LocalizationManager.swift # Runtime language switching
│ ├── Models/ # Core Data entities
│ └── Tests/ # Unit, integration & stress tests
├── resources/
│ ├── heuristics/ # 14 JSON extraction rule files
│ └── documentation/ # Internal guides
└── project.yml # XcodeGen configuration
| Component | Technology |
|---|---|
| UI | SwiftUI |
| OCR | Vision Framework |
| AI Model | Qwen2.5-0.5B-Instruct (4-bit, MLX) |
| Inference | MLX + llama.xcframework |
| Persistence | Core Data (SQLite) |
| Sync | iCloud Documents |
| Build System | XcodeGen + Xcode |
| Metric | Target | Actual |
|---|---|---|
| OCR Processing | < 2s | 1.8s avg |
| AI Inference | < 300ms | 250ms avg |
| Memory Usage | < 150MB | 140MB peak |
| Extraction Accuracy | > 90% | 95%+ |
| Cache Hit Rate | > 60% | 78% |
| Crash Rate | 0% | 0 / 1000 ops |
- 100 consecutive inferences: passed (avg 380ms)
- 200 receipt validation: 96% success rate
- Memory leak test: no leaks detected
- 1000 operations: zero crashes
- 100% on-device processing — OCR and AI inference run locally
- No cloud dependencies — Works fully offline after installation
- No analytics or telemetry — Zero data collection
- Core Data encryption available for sensitive data
- Camera and Photo Library permissions required for scanning
# Run unit tests
xcodebuild test -project PrivExpensIA.xcodeproj \
-scheme PrivExpensIA \
-sdk iphonesimulator \
-destination 'platform=iOS Simulator,name=iPhone 15'Test coverage: OCR 85% | Core Data 80% | UI 75% | Overall 80%+
MIT License. See LICENSE for details.
Built by Mr D — part of the Infinity Cloud ecosystem.