Fraud Triage Agent
shivamdhawani24-bit/fraud-triage-agent
Triagem automática de fraude com raciocínio agentic e integração de regras
Ver no GitHubSistema agentic que analisa alertas de transações fraudulentas usando a Claude tool-use API para realizar triagem inteligente de sinais de fraude. Integra raciocínio estruturado com mecanismo de regras customizável, gerando decisões (APPROVE/REVIEW/BLOCK) com score de confiança e narrativa de justificativa para operações de fraude.
Instalação
npx convez-ai shivamdhawani24-bit/fraud-triage-agent1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar análise L1 de alertas de transações em tempo real
- Reduzir carga de analistas de fraude com decisões estruturadas
- Integrar Claude com regras de negócio existentes (Drools, SQL, custom engines)
Veredito

Implementação prática de agentic AI para fraude ops. Demonstra bem tool-use do Claude com integração real de regras, ideal para times de compliance que precisam escalar triagem manual. Código bem estruturado e documentado, com exemplos claros de detecção de ATO, card testing e anomalias de velocidade.
README
A tradução para o português deste README ainda está na fila.
🛡️ Fraud Signal Triage Agent
An agentic AI system that ingests raw transaction alerts, reasons over fraud signals using Claude's tool-use API, and outputs a structured triage decision with confidence scoring and a natural-language reasoning narrative.
Built to demonstrate how agentic AI can reduce L1 analyst load in fraud ops by automating structured triage reasoning at scale.
Demo
How It Works
Transaction Alert (JSON)
↓
Claude Agent (claude-opus-4-5)
↓
[Tool Call] lookup_fraud_rules(signal_type)
↓
Mock Rules Engine (rules.py)
↓
Structured Decision: APPROVE / REVIEW / BLOCK
+ Confidence Score + Reasoning Narrative
The agent uses Claude tool use — not just prompt engineering. It actively calls a rule-lookup tool mid-reasoning, simulating how a real fraud ops system integrates an AI layer with an existing rules engine (Drools, custom SQL rules table, etc.).
Fraud Signals Covered
| Signal Type | Rule ID | Description |
|---|---|---|
| ATO | R-001 | Account Takeover — credential changes + device mismatch |
| Card Testing | R-002 | Enumeration pattern — micro-txns + high decline rate |
| Synthetic Identity | R-003 | Bust-out risk — thin file + new account |
| Velocity Anomaly | R-004 | Spend/txn velocity spike vs. account baseline |
| Device Risk | R-005 | Unrecognized device, emulator, VPN/proxy signals |
| New Account Risk | R-006 | Early life fraud — no behavioral baseline |
Tech Stack
- Anthropic Claude —
claude-opus-4-5with tool use - Python 3.10+
- Streamlit — interactive UI
- Custom rules engine — simulates production fraud rule DB
Getting Started
1. Clone the repo
git clone https://github.com/shivamdhawani24-bit/fraud-triage-agent.git
cd fraud-triage-agent
2. Install dependencies
pip install -r requirements.txt
3. Set your Anthropic API key
export ANTHROPIC_API_KEY=your_api_key_here
Get a free API key at console.anthropic.com
4. Run the Streamlit app
streamlit run app.py
5. Or run the agent directly
python agent.py
Sample Output
{
"decision": {
"decision": "BLOCK",
"confidence": 94,
"fraud_type": "Account Takeover (ATO)",
"signals_detected": [
"password_changed_last_24h",
"email_changed_last_48h",
"device_mismatch",
"ip_country_mismatch",
"high_velocity_new_account"
],
"reasoning": "This transaction exhibits 5 concurrent ATO indicators including credential changes, device mismatch, and a foreign IP on a 3-day-old account with 8 transactions in 24 hours. The combination of account mutation signals with behavioral velocity anomaly on a new account is a high-confidence ATO pattern.",
"recommended_action": "Block transaction immediately. Lock account and trigger step-up verification. Flag for manual review by fraud ops team."
},
"reasoning_trace": [
"Checking rules for signal: ATO",
"Checking rules for signal: velocity"
],
"alert_id": "TXN-20240607-001"
}
Project Structure
fraud-triage-agent/
├── agent.py # Claude agentic logic with tool use
├── rules.py # Mock fraud rules engine
├── app.py # Streamlit UI
├── requirements.txt
└── sample_alerts/
└── test_cases.json # 5 sample alerts for testing
About
Built by Shivam Dhawan — Senior Fraud & Risk Analyst with 12+ years at Intuit, Chime, and Airbnb.
This project draws on real fraud typologies from production systems — ATO detection, card fraud rings, synthetic identity, and payments risk — translated into an agentic AI architecture.