GargiG15/Accounts_Payable_AI_Auditor_

AP Workflow OpenEnv 🧾

  

Overview

AP Workflow OpenEnv simulates the Accounts Payable (AP) invoice processing workflow — a critical real-world finance task performed by every company on earth. AP clerks handle incoming invoices by:

1. 1.Validating invoice data against Purchase Orders (POs)
2. 2.Detecting fraudulent or duplicate invoices
3. 3.Routing high-value invoices for manager approval
4. 4.Posting approved invoices to the General Ledger (GL)

Humans spend 20–40 hours/week on AP tasks. AI agents that can automate this accurately save enterprises millions annually. This environment enables RL researchers and agent developers to train and evaluate agents on this domain.

Why This Environment?

• —Real-world utility: AP processing is a billion-dollar automation target. No existing OpenEnv covers financial document workflows.
• —Multi-step reasoning: Requires reading structured data, cross-referencing POs, detecting anomalies, and making routing decisions.
• —Fraud detection: Agents must identify signals like amount spikes, fake vendor names, and round-number patterns.
• —Partial rewards: Dense reward shaping guides agents step-by-step rather than sparse end-of-episode signals.

Action Space

Actions are structured JSON with the following tools:

Example Action

{
  "tool_call": "validate_data",
  "parameters": {
    "invoice_id": "INV-A1B2C3D4",
    "amount": 1250.00,
    "vendor_id": "V002"
  },
  "reasoning": "Checking if invoice amount matches PO approved amount within 5% tolerance"
}

Observation Space

Each observation contains:

Invoice Object

{
  "invoice_id": "INV-A1B2C3D4",
  "vendor_name": "TechParts Ltd.",
  "vendor_id": "V002",
  "amount": 1250.00,
  "date": "2025-01-15",
  "due_date": "2025-02-14",
  "po_number": "PO-XYZ123",
  "line_items": [...],
  "status": "pending",
  "is_duplicate": false,
  "is_fraudulent": false
}

Tasks

Easy: Single Invoice Validation

• —Max Steps: 20
• —Description: Validate one basic invoice against a Purchase Order. Check amount, date, vendor, and post to the correct GL account.
• —Grader:
• —1.0 = Correctly validated + posted to correct GL
• —0.7 = Posted but wrong GL code
• —0.5 = Validated but not posted
• —0.0 = Valid invoice rejected
• —Expected baseline (GPT-4o): ~0.92

Medium: Fraud & Duplicate Detection

• —Max Steps: 35
• —Description: Process 3 invoices: 1 valid, 1 fraudulent, 1 duplicate. Correctly classify each.
• —Grader:
• —(TP_fraud + TP_dup + TP_valid_posted) / total + 0.2 × prevented_loss_ratio
• —False positives penalized at -0.1 each
• —Expected baseline (GPT-4o): ~0.67

Hard: Full AP Cycle with Approvals

• —Max Steps: 50
• —Description: Process 5 invoices over a simulated 3-day cycle. Includes 1 high-value invoice requiring approval routing, 1 fraudulent, 1 duplicate. Must post all valid invoices to correct GL accounts.
• —Grader:
• —0.4 × validation_accuracy + 0.3 × approval_routing_correct + 0.3 × ledger_accuracy
• —Episode ends when all invoices processed or max steps reached
• —Expected baseline (GPT-4o): ~0.41

Reward Function

Rewards are shaped to provide dense feedback throughout the episode:

Project Structure

ap-openenv/
├── openenv.yaml     # OpenEnv metadata and task list
├── models.py        # Pydantic: Observation, Action, Reward, StepResult
├── env.py           # APWorkflowEnv: step() / reset() / state()
├── grader.py        # Task-specific graders (0.0–1.0)
├── data_generator.py # Invoice/PO generation with fraud injection
├── server.py        # FastAPI OpenEnv endpoints
├── inference.py     # Baseline inference script (OpenAI client)
├── Dockerfile       # Container definition
├── requirements.txt # Python dependencies
└── README.md        # This file

Setup & Usage

Local Development

# Install dependencies
pip install -r requirements.txt

# Start the server
python server.py
# → Server running at http://localhost:8000

# In another terminal, run baseline
export OPENAI_API_KEY=sk-your-key
python inference.py

Docker

# Build
docker build -t ap-env .

# Run
docker run -p 8000:8000 ap-env

# Run baseline against dockerized env
export OPENAI_API_KEY=sk-your-key
export AP_ENV_URL=http://localhost:8000
python inference.py

Hugging Face Spaces

Deploy directly to HF Spaces — the server runs FastAPI on port 7860 (set PORT=7860).

API Reference

POST /reset

{"task_id": "single_invoice_validation", "seed": 42}

POST /step

{
  "tool_call": "validate_data",
  "parameters": {"invoice_id": "INV-...", "amount": 1000.0},
  "reasoning": "Checking fields"
}

GET /state

Returns full internal state for debugging.

POST /grade

{"task_id": "single_invoice_validation"}

Returns {"score": 0.92, "breakdown": {...}, "passed": true, "feedback": "..."}

Baseline Scores

Reproduce with:

export OPENAI_API_KEY=your-key
export MODEL_NAME=gpt-4o-mini
python inference.py

Environment Variables

License

MIT License — see LICENSE.