Berlin AI Engineering Group by m|rig

Companies in e-commerce and digital payments are under constant pressure to detect fraud early, reduce financial losses, and minimize friction for legitimate customers.

Traditional fraud detection systems — rules engines and black-box ML models — can be effective at flagging suspicious transactions, but they often fail where it matters most:

• Why was a transaction flagged?
• Can a fraud analyst trust and understand the decision?
• How do you explain fraud risk to business teams or customers?

This is where Agentic AI and Large Language Models (LLMs) unlock a new layer of value.
But moving from a fraud score to a clear, human-readable explanation of fraudster behavior requires more than just adding an LLM on top of a model.

How do you design an Agentic AI system that not only detects potentially fraudulent transactions, but can also reason about the signals, patterns, and behaviors involved — and explain them in plain language?

In this 60-minute interactive meetup, we’ll walk through the end-to-end design and implementation of an explainable fraud detection system, combining classical fraud detection logic with LLM-based reasoning and explanation.

The session will include a live, code-driven demonstration using Python and LangChain, with concrete e-commerce fraud examples, showing how Agentic AI can support fraud analysts, risk teams, and decision-makers.

### You’ll learn

🔍 Core Concepts of Transaction Fraud Detection
How fraud typically manifests in e-commerce:

• Common fraud patterns and behaviors
• Signals used in transaction-level detection
• Limitations of rules-based and black-box models

🏗️ Designing an Agentic AI Fraud System
How to move from detection to explanation:

• Separating detection, reasoning, and explanation
• Agent roles (scoring, investigation, explanation)
• Guardrails and consistency in risk explanations
• Human-in-the-loop design for fraud analysts

🤖 Using LLMs to Explain Fraudster Behavior
How LLMs can:

• Analyze transaction context and signals
• Reason about suspicious behavioral patterns
• Generate clear, plain-text explanations
• Adapt explanations to different audiences (analyst vs business)

🐍 Live Demo: Building an Explainable Fraud Agent with LangChain & Python
A practical walkthrough including:

• Transaction data ingestion
• Fraud signal aggregation
• Agent orchestration with LangChain
• Step-by-step reasoning traces
• Plain-language fraud explanations for e-commerce transactions

🧪 Interactive Q&A and Architecture Discussion
Discuss:

• Trust and explainability in fraud systems
• Risks of hallucination and how to control them
• Regulatory and audit considerations
• Where Agentic AI fits into existing fraud stacks

📅 Duration: 60 minutes

🛠️ Tech Stack: Python, LangChain, LLMs, Agentic AI
URL: https://events.teams.microsoft.com/event/6f6dc1ed-6256-4642-8d73-e37b48767c3b@d94ea0cb-fd25-43ad-bf69-8d9e42e4d175