Data & Machine LearningSTATUS: Open Source 1 0

LedgerQuarry Document Parser

Multimodal vision and OCR financial document extraction pipeline

LedgerQuarry replaces brittle coordinate-based template OCR with flexible multimodal vision models. It accepts scanned receipts, PDFs, and invoices of any layout, extracting structured financial line items, dates, and amounts into validated accounting schemas with per-field confidence metrics.

Inspect GitHub Source Code

01 // THE BUSINESS PROBLEM & SITUATION

Operational Context

Finance and accounting departments waste hundreds of hours manually entering vendor invoice line items into ERP software.

CLIENT / ENVIRONMENT: Created for finance teams processing diverse, unstructured supplier invoices and receipts.

02 // SYSTEM ARCHITECTURE & SOLUTION

What Was Built

Constructed an automated processing pipeline utilizing Vision AI models coupled with strict schema validation to extract structured financial data.

LEDGERQUARRY // MULTIMODAL_OCRDocument Extraction Pipeline

UPLOAD

PDF / Receipt

VISION AI

Multimodal Model

VALIDATE

Zod Schema

ERP SYNC

Typed JSON

SYSTEMS_ARCHITECTURE_EXPLODER // LAYERED_STACKFULL-STACK SYSTEM DECOUPLING

Frontend Layer

Next.js 16 App Router, React 19, Tailwind CSS, TypeScript

API & Middleware

FastAPI, Pydantic Schema Validation, Rate Limiters

Orchestration

LangGraph Multi-Agent Supervisor & State Machines

Model Providers

OpenAI GPT-4o, Anthropic Claude 3.5, Vision AI

Data Stores

PostgreSQL (Prisma), Qdrant Vector Store, Redis Caching

Observability

LangSmith traces, Pydantic audit logs, Error retries

03 // ENGINEERING CONTRIBUTION & FEATURES

Exact Contribution

Built secure document upload and processing API with Express/Node.js
Engineered structured vision prompts for zero-shot invoice parsing
Implemented per-field confidence scoring and automated anomaly detection
Created ERP export formatters (QuickBooks, Xero, CSV)

Key Capabilities Built

✓ Multi-Page PDF & Image Document Parsing

✓ Granular Line-Item and Tax Breakdown Extraction

✓ Per-Field Confidence Scoring & Validation

✓ Automated Currency & Date Normalization

✓ ERP Integration Webhooks

DATA_PROVENANCE_TIMELINE // AUDIT_TRACETRACE_ID: #TR-992041

14:22:01.002Source Ingestion

Received webhook payloadOK

14:22:01.045Schema Validation

Pydantic structured field checkOK

14:22:01.210LangGraph Reasoning

Evaluated criteria rulesOK

14:22:01.350Human Audit Check

Confidence threshold evaluationPASSED

14:22:01.480Database Commit

Created audit record in PostgreSQLCOMMITTED

*Illustrative system trace — demonstration data, not a client result.

04 // EXECUTION PIPELINE

Step-by-Step Workflow

STEP 01User or automated webhook uploads invoice document.

STEP 02Vision model analyzes visual structure and text hierarchy.

STEP 03Schema validator inspects extracted line items, subtotals, and vendor data.

STEP 04Confidence scoring flags low-certainty fields for human review.

STEP 05Validated structured JSON dispatches to accounting software endpoint.

05 // ENGINEERING LESSONS & OUTCOMES

Technical Challenges Overcome

• Parsing low-contrast mobile receipt photos with glare and folds.
• Resolving multi-currency line items and varying international tax conventions.

Measurable Outcomes

→ Achieved 95%+ extraction accuracy across varied vendor invoice formats in test sets.

→ Eliminated reliance on rigid coordinate-based OCR templates.

Lessons & Engineering Rules

"Multimodal LLMs provide vastly superior generalization over traditional coordinate OCR for diverse document layouts."
"Mathematical cross-validation (verifying line items sum to the subtotal) catches extraction hallucinations automatically."

NEXT RELATED CASE STUDY

Orchestrion Multi-Agent Studio

View Study

PROJECT METADATA

ARSLAN'S ROLEMachine Learning Engineer

YEAR & STATUS2024 // Open Source

CATEGORYData & Machine Learning

TECHNOLOGY STACK

JavaScriptNode.jsVision AIExpressOCRPydantic/Zod

GITHUB METRICS

Stars: 1

Forks: 0

Verified GitHub Update: 8/15/2026

Need a similar solution?

If your business faces similar data, automation, or software bottlenecks, let's discuss your requirements.

Start a Conversation