PaymentGuard: FinTech & Payment Solutions

Octopus Hackathon 2026

Multi-agent system for PCI-DSS safe synthetic payment data and fraud detection.

---

Core features

1. Few-shot synthetic fraud data generation
   Uses prompt-based context-learning: the LLM sees a small set of real (and optionally synthetic) transaction rows and generates new rows that match the same distribution. The pipeline enforces schema alignment, value ranges, and PCI-DSS safety (no real PAN/CVV). Class-conditional generation and filtering keep fraud/non-fraud balance controllable.

2. Explainable fraud detection
   Few-shot LLM prediction outputs Fraud (0/1), a score, and a Reason (2–3 sentences of judgment logic) per transaction. Every decision is interpretable: what stands out, how it relates to fraud/legitimate behavior, and a clear conclusion. The web app displays Fraud, Fraud_Score, and Reason in the results table.

3. Multi-agent pipeline
   FinTech Payment Agent (schema + PCI-DSS), Privacy Guard (PII/PCI detection and protection), Synthetic Pipeline V7 (meta-optimization, class-conditional generation, filter and rank), and Impact Evaluator (business impact and reports). One run produces PCI-safe synthetic data and an HTML report; the web app then uses original + synthetic data as few-shot sources for the fraud predictor.

---

Quick run

From project root:

• Windows: python octopus\run_payment_guard.py --fast
• macOS / Linux: python octopus/run_payment_guard.py --fast

Opens octopus/paymentguard_report.html after run. See RUN_DEMO.md for details. please make sure that you need to fill a togetherai api key in the demo_payment_guard.DEFAULT_TOGETHER_API_KEY

---

Features

• FinTech Payment Agent — Schema detection, PCI-DSS constraints, payment scoring
• Privacy Guard Agent — PII/PCI detection and protection
• Synthetic Pipeline V7 — Class-conditional generation and filtering
• Impact Evaluator — Business impact and social value

Demo splits data into train/test (80/20). Train set drives synthesis and few-shot; test set is saved as octopus/data/test_transactions.csv (10 rows) for the web app.

---

Web app: LLM fraud prediction

1. Run the demo once to generate last_original.csv and last_synthetic.csv.
2. Start: python -m octopus.app_fraud_detection
3. Open http://127.0.0.1:5000 and upload octopus/data/test_transactions.csv.

Predictions show Fraud (0/1), Fraud_Score, and Reason (judgment logic) on separate rows.

---

Optional commands

Verify synthetic utility (A vs B):

python -m octopus.verify_synthetic_utility --no-api
python -m octopus.verify_synthetic_utility

Distribution plot (PCA 2D):

python -m octopus.visualize_distribution --no-api

Output: octopus/distribution_plot.png.

---

Dependencies

pip install -r octopus/requirements.txt

Python 3.8+, pandas, numpy, scikit-learn, openai. Set TOGETHER_API_KEY to override bundled key.

---

Project layout (submission)

octopus/
├── README.md
├── RUN_DEMO.md
├── requirements.txt
├── demo_payment_guard.py
├── run_payment_guard.py
├── run_payment_guard.sh
├── payment_guard_system.py
├── synthetic_pipeline_v7.py
├── cllm/src/llm_gen.py
├── fintech_payment_agent.py
├── privacy_guard_agent.py
├── impact_evaluator_agent.py
├── domain_adapter_agent.py
├── task_coordinator_agent.py
├── report_ui.py
├── fraud_detection_agent.py
├── llm_fraud_predictor.py
├── app_fraud_detection.py
├── run_fraud_web.py
├── verify_synthetic_utility.py
├── visualize_distribution.py
├── load_fraud_data.py
├── THE_STORY_FINTECH.md
├── PITCH_SCRIPT_FINTECH.md
└── SUBMISSION_FINTECH_7PTS.md