An open-source AR/AP automation engine that ingests invoices via OCR and Vision LLMs, reconciles payments, forecasts DSO, detects fraud, orchestrates dunning, syncs with external accounting platforms, and handles external payments.
APEX is an open-source accounts payable and receivable automation engine. It processes invoices through an OCR and Vision AI pipeline across multiple channels (email, messaging), reconciles them against payment records, projects Days Sales Outstanding (DSO) with time-series forecasting, detects financial anomalies using Isolation Forest and Semantic LLM validation, and orchestrates dunning communications. It features bidirectional sync with external accounting platforms and native rails for external payments.
Extracts structured data from invoices and scanned documents using Tesseract OCR and Vision LLMs. Supports multi-channel ingestion (webhooks, email, messaging). Validates fields (amount, date, vendor, line items), detects near-duplicate invoices, and writes validated records to a local SQLite database.
Input: Invoices via multi-channel (PDFs, images, messaging)
Output: Structured invoice records — validated, deduplicated, persisted
Runs automatic payment matching against invoice records using configurable tolerance logic. Projects DSO (Days Sales Outstanding) with Facebook Prophet time-series models. Detects financial fraud and anomalous payment patterns via Isolation Forest. Generates aging reports and cash collection analytics.
Input: Invoice database + payment records
Output: DSO forecast · fraud flags · reconciled AR/AP status · aging report
LangGraph orchestrates specialized agents across the AR/AP lifecycle, including a Virtual AI CFO for cash flow insights and a Semantic Approval agent for intelligent routing. Generates dunning email drafts. Syncs processed financial outcomes bidirectionally with external accounting platforms and executes transfers via external payment platforms.
Input: Reconciled records + DSO + semantic signals
Output: Dunning emails · CFO insights · external payments · accounting sync
graph TD
A["PDF Invoices / Scanned Documents"] -->|OCR pipeline| B["OCR Extraction Agent<br/>app/agents/ocr_extraction.py<br/>Tesseract · pytesseract"]
B -->|Structured invoice data| C[("SQLite Database<br/>apex_ar.db")]
C -->|Invoice records| D["Matching & Reconciliation Agent<br/>app/agents/matching_reconcile.py<br/>RapidFuzz"]
C -->|Payment history| E["DSO Forecast Agent<br/>app/agents/dso_forecast.py<br/>Prophet"]
C -->|Transaction patterns| F["Fraud Detection Agent<br/>app/agents/fraud_detection.py<br/>Isolation Forest"]
D -->|Reconciled status| G["LangGraph Orchestrator<br/>app/agents/orchestrator.py"]
E -->|DSO projections| G
F -->|Anomaly flags| G
G -->|Approved actions| H["Dunning Communication Agent<br/>app/agents/dunning_communication.py<br/>Claude LLM / Offline"]
G -->|Outcomes| I["AFIS Sync Agent<br/>app/agents/afis_sync.py"]
G -->|REST API| J["FastAPI Backend<br/>app/main.py · Port 8002"]
J -->|/api/v1| K["Glassmorphic Dashboard<br/>frontend/ · Chart.js"]
| Endpoint | Method | Description |
|---|---|---|
/api/v1/invoices |
GET / POST |
List invoices or upload new invoice for OCR processing |
/api/v1/reconciliation |
GET / POST |
Trigger payment matching and view reconciliation status |
/api/v1/dunning |
GET / POST |
List pending dunning actions or trigger communication |
/api/v1/fraud |
GET |
Fraud alerts and anomaly detection results |
/api/v1/forecast |
GET |
DSO projections with time-series confidence bands |
/api/v1/reports |
GET |
Aging reports and AR/AP cash flow summaries |
/api/v1/system |
GET |
System status, AI mode (llm or offline), version |
/api/v1/payments |
POST |
Gateway for external payments and scheduling |
/api/v1/sync |
GET / POST |
Bidirectional sync with external platforms |
/api/v1/webhooks |
POST |
Omnichannel ingestion (messaging/email) |
| Component | Technology |
|---|---|
| Backend | FastAPI 0.115 (Python 3.11+) |
| Agent Orchestration | LangGraph 0.2.39 · LangChain 0.3.7 |
| OCR | Tesseract · pytesseract · pdf2image · Pillow |
| DSO Forecasting | Prophet 1.1.5 |
| Fraud Detection | scikit-learn · Isolation Forest |
| Fuzzy Matching | RapidFuzz |
| AI Narrative | Anthropic Claude (optional) · offline heuristics fallback |
| Database | SQLite (zero-server, local-first) · SQLAlchemy |
| Dashboard | HTML + CSS + JavaScript · Chart.js |
| aiosmtplib (async dunning dispatch) | |
| Testing | pytest · pytest-asyncio · httpx |
Local-first, privacy by design. All invoice documents and payment records stay on the SME's machine.
Semantic Approvals & Virtual CFO. APEX utilizes LLMs not just for data extraction, but to semantically evaluate if an invoice requires human review, and provides a virtual AI CFO that outputs plain-text cash flow recommendations.
External Platform Integration. Seamlessly integrates with external accounting platforms for bidirectional ledger syncing, and external payment platforms for one-click payment execution.
Offline-first dunning. Even without Claude, APEX generates structured dunning communications using rule-based templates.
NIST AI RMF 1.0 alignment. Every agent action is logged to a persistent audit trail following NIST governance principles.
APEX is built for SME finance teams, accountants, and AR/AP managers who need automation without enterprise-grade ERP costs.
You do not need a data science background. Upload your invoices, connect to your payment records, and run python run.py. APEX handles OCR, matching, forecasting, and dunning automatically.
git clone https://github.com/afild/APEX.git
pip install -r requirements.txt
python run.pyOpen http://localhost:8002 in your browser.
LLM Mode — set environment variable and restart:
# Linux/macOS
export ANTHROPIC_API_KEY=your_key_here
# Windows
set ANTHROPIC_API_KEY=your_key_here
python run.pyOffline Mode (default): rule-based dunning templates and deterministic reconciliation logic. Full OCR, forecasting, and fraud detection operate identically in both modes.
- Python 3.11 or higher
- Git
- Tesseract OCR binary (install guide)
# 1. Clone and navigate
git clone https://github.com/afild/APEX.git
# 2. Create virtual environment
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
# 3. Install dependencies
pip install -r requirements.txt
# 4. Configure environment
cp .env.example .env
# Set TESSERACT_CMD path and optionally ANTHROPIC_API_KEY in .env
# 5. Launch
python run.pypytest tests/ -v| NIST Function | APEX Implementation |
|---|---|
| GOVERN | MIT License · open audit trail · traceable agent decision logic |
| MAP | AR/AP domain scoped to SME invoice processing · documented assumptions |
| MEASURE | Automated pytest suite · DSO residuals computed per run · fraud scoring per transaction |
| MANAGE | Offline fallback · duplicate flagging · human-review step before dunning dispatch · explainable Isolation Forest |
The AI dunning agent sends only structured invoice context (amount, days overdue, vendor name) to the LLM API — never raw invoice images or scanned documents.
APEX/
├── app/
│ ├── main.py ← FastAPI app · route registration · static serving
│ ├── config.py ← Pydantic settings · environment variables
│ ├── llm_client.py ← Provider-agnostic LLM client (Claude + offline)
│ ├── agents/
│ │ ├── orchestrator.py ← LangGraph state machine · agent coordination
│ │ ├── ocr_extraction.py ← Tesseract OCR · field parsing · validation
│ │ ├── matching_reconcile.py ← RapidFuzz payment matching · reconciliation logic
│ │ ├── dso_forecast.py ← Prophet time-series DSO forecasting
│ │ ├── fraud_detection.py ← Isolation Forest anomaly scoring
│ │ ├── dunning_communication.py ← Dunning email generation (LLM + offline)
│ │ └── afis_sync.py ← Financial outcome sync to AFIS layer
│ ├── api/
│ │ ├── router.py ← API router aggregation (prefix /api/v1)
│ │ ├── invoices.py ← Invoice CRUD and OCR trigger
│ │ ├── reconciliation.py ← Payment matching endpoints
│ │ ├── dunning.py ← Dunning action endpoints
│ │ ├── fraud.py ← Fraud alert endpoints
│ │ ├── forecast.py ← DSO forecast endpoints
│ │ ├── reports.py ← Aging report and summary endpoints
│ │ └── system.py ← Health check and system status
│ ├── database/
│ │ ├── db_manager.py ← SQLite init · schema creation
│ │ ├── models.py ← SQLAlchemy ORM models
│ │ └── schema.sql ← Tables: invoices · payments · audit_log
│ ├── ml/
│ │ ├── dso_prophet.py ← Prophet model training and forecasting
│ │ └── isolation_forest.py ← Isolation Forest fraud model
│ ├── plugins/
│ │ ├── afis_reader.py ← Read-only AFIS database connector
│ │ ├── axis_reader.py ← External data integration
│ │ ├── email_connector.py ← aiosmtplib async email dispatch
│ │ └── pdf_storage.py ← Local PDF storage management
│ └── skills/
│ ├── extract_invoice_fields.py
│ ├── detect_duplicate_invoice.py
│ ├── match_invoice_payment.py
│ └── generate_dunning_email.py
├── data/ ← Sample invoice datasets
├── docs/
│ └── images/ ← Architecture diagrams
├── frontend/
│ ├── index.html ← Glassmorphic dark-mode dashboard
│ ├── styles.css
│ └── app.js ← Chart.js · API integration · dunning UI
├── tests/
│ └── test_api.py ← API endpoint tests
├── .env.example ← Environment variable template
├── CHANGELOG.md
├── requirements.txt
└── run.py ← Single-command launcher
Areas where contributions are most needed:
- Additional OCR backends (AWS Textract, Azure Form Recognizer)
- Support for DOCX and CSV invoice formats
- Multi-currency reconciliation support
- Docker Compose setup for zero-dependency deployment
- Full AR/AP automation pipeline with Vision OCR, reconciliation, Semantic Approvals, AI CFO, and external integrations
MIT License — free to use, adapt, and redistribute.