跳到正文
DEV Community· Cristian Núñez (DiGitax)·· 2 小時前AI 評分34

DiGitax Suite 如何以 Gemini AI 與 Vertex AI 雙引擎架構，在 4.5 分鐘處理 600 張真實税務發票

How We Processed 600 Real Tax Invoices in 4.5 Minutes Using a Dual-Engine Gemini AI & Vertex AI Architecture

AI 導讀

DiGitax Suite 以雙引擎 AI 文件處理流程，在 4 分 31 秒內處理 600 張真實發票，欄位擷取準確率達 100%。系統以 Gemini AI 支援免費日常處理，並以 Vertex AI 處理月底批量結算，每份文件收取 6 Credits。

正文

Cover image for How We Processed 600 Real Tax Invoices in 4.5 Minutes Using a Dual-Engine Gemini AI & Vertex AI Architecture

Cristian Núñez (DiGitax)

Manual data entry remains the single biggest bottleneck for Certified Public Accountants (CPAs) and SMEs in Latin America and the Caribbean. In the Dominican Republic, preparing the monthly DGII Form 606 (Purchases), Form 607 (Sales), and IT-1 (VAT/ITBIS) declarations for a mid-sized business with 600 monthly receipts typically consumes 15 hours of manual typing and tax ID validation.

At DiGitax Suite (Corporate Profile — DiGitax LLC), we engineered a dual-engine AI document processing pipeline that cuts those 15 hours down to 4 minutes and 31 seconds with 100% field extraction accuracy—while keeping daily AI ingestion free for accountants.

Here is the architecture and the full benchmark data from our 600-Invoice AI Case Study (Gemini AI vs. Vertex AI).

1. The Problem with Single-Tier OCR SaaS

Most accounting scanners fail in production for two reasons:

  1. Expiring Monthly Quotas: Traditional OCR tools charge per seat and wipe out unused document credits at the end of every month.
  2. Zero Fiscal Filtering: Employees send everything via WhatsApp or email—thermal receipts, quotes/proformas, delivery notes (conduces), consumer receipts without tax credit (B02), and duplicate PDFs. If an OCR tool blindly dumps all of that into the ledger, the accountant spends more time cleaning up the mess than typing it manually.

2. Our Solution: Dual-Engine AI Routing (0 Credits vs. 6 Credits/doc)

Instead of forcing every document through an expensive enterprise endpoint, DiGitax routes documents across two specialized engines backed by a Unified Non-Expiring Prepaid Credit Wallet:

Metric / Capability Track 1: Google Gemini AI (0 Credits / $0.00 Free) Track 2: Google Cloud Vertex AI (6 Credits / doc)
Target Workload Continuous daily ingestion (WhatsApp, mobile app, email) End-of-month bulk closures & multi-page PDF batches
Throughput Speed 45 documents / hour (+20 free scans/day per Google key) 1,500 to 4,000 documents / hour (~4.5 min per 300 docs)
Multi-Invoice PDF Splitting Up to 3 invoices per PDF Unlimited (splits 50–100 page stapled PDFs automatically)
Extraction Accuracy (RNC, NCF, ITBIS) 100% 100%
Cost per Document $0.00 USD (0 Credits) $0.0042 – $0.0060 USD (6 Credits)

3. The 6-State Fiscal Tray: Declarados vs. Descartados

Before any extracted document touches the General Ledger, it passes through our 4-Filter DGII Validation & Depuration Pipeline:

  1. Fiscal & Accounting Track (Left Rail):

    • Procesados → Contabilizados → Declarados
    • Verifies live taxpayer status (RNC Activo) and voucher sequence (NCF Serie B / e-CF Serie E under Law 32-23) against the DGII Official Registry.
    • Generates an atomic double-entry journal entry ($\text{Debits} = \text{Credits}$) and locks the period with an immutable cryptographic lock once declared.

  2. Control & Depuration Track (Right Rail):

    • Pendientes → Incidencias → Descartados
    • Quotes, proformas, non-fiscal receipts, consumer invoices (B02), and exact [RNC + NCF] duplicates are automatically diverted to Descartados with an explicit audit tag—never polluting the DGII Form 606, 607, 608, 609, or 623.

4. Explore the Full Benchmark & Free Tools

來源：DEV Community · dev.to

#教學/實踐#部署/工程#Google