Innobi Fiscra : Native iOS Receipt & Expense Intelligence

Case Study: Innobi Fiscra — Native Receipt-to-Record Pipeline

Receipt capture is deceptively hard: vendors print inconsistently, tax formats vary, and OCR confidence can drop quickly on tilted, wrinkled, or low-contrast scans. For a finance app, the real problem is not “extracting text” — it is producing a trustworthy, user-verifiable record that can stand up in real expense tracking workflows.

Innobi Fiscra addresses this with a pragmatic approach: on-device Vision OCR as the fast default, AI-assisted structuring when needed, and a clear confirmation UI before anything is saved. The result is a clean pipeline from receipt photo → extracted fields → categorized expense entry → searchable archive, built as a portfolio-grade native iOS product.

The Challenge

• Inconsistent receipt layouts: Vendor formats vary widely, making naive regex parsing unreliable.
• Accuracy vs. speed tradeoff: Users want instant results, but finance records require verification and consistency.
• Trust & explainability: Any AI-assisted extraction must be reviewable and editable before saving.

The Solution & Architecture

• Tier 1 (Local OCR): Apple Vision text recognition extracts raw text quickly on-device.
• Tier 2 (AI Structuring): Optional AI step converts OCR text into structured fields (vendor, date, subtotal, tax, total, receipt number) with a confidence signal.
• Tier 3 (User Confirmation): Review-first UI lets users edit fields and confirm before committing to storage.
• Offline-first capture: Photos and draft records can be created without network; enrichment can run later.
• Searchable archive: Receipts become normalized records that can be filtered and exported as summaries.

Swift Parsing Model (Receipt Fields)

A small domain model keeps OCR and AI results aligned with an explicit “user-confirmed” state.

import Foundation

struct ReceiptRecord: Identifiable, Codable {
  let id: UUID
  var vendorName: String
  var vendorAddress: String?
  var transactionDate: Date?
  var receiptNumber: String?

  var subtotal: Decimal?
  var taxAmount: Decimal?
  var totalAmount: Decimal?

  // Raw OCR text retained for explainability / re-parsing
  var ocrText: String

  // Pipeline signals
  var extractionTier: ExtractionTier
  var confidenceScore: Double // 0.0 - 1.0
  var isUserConfirmed: Bool
}

enum ExtractionTier: String, Codable {
  case tier1VisionOCR
  case tier2AIStructuring
  case manualOnly
}

This model supports a transparent pipeline: store OCR text, surface confidence, and require confirmation before finalizing the record.