Fraud detection and credit line optimization
Two parallel tracks with the same partner — a data-driven credit-limit policy and an LLM-based fraud detection agent for real-time B2B payment risk.
ReLU worked with Two on two related projects during Fall 2025 — one on credit policy, one on fraud detection.
Optimizing credit line policy
ReLU helped Two develop a data-driven method for setting personalised credit limits that maximise business value while keeping portfolio risk under control. The goal was to recommend the right limit for each buyer based on their financial profile, behaviour, and Two’s strategic objectives, all within the company’s global risk constraints.
The team started by analysing Two’s historical data to understand data quality and the behavioural patterns that drive credit usage and emerging risk. Together with Two’s Data Science team, they explored several modelling and optimisation approaches before converging on a framework that dynamically adjusts limits as customers become more or less risky. The semester ended with an implemented and validated solution, designed to responsively protect capital at risk while giving reliable buyers room to grow.
Automating fraud detection with LLMs
ReLU also helped Two develop an AI-driven fraud detection system that automates risk assessments in real time, reducing reliance on manual review while improving both speed and accuracy. The goal was to identify fraudulent orders — ranging from third-party impersonation to first-party abuse from scam companies — by combining internal transaction data with external signals about the buyer and its representatives, enabling faster decisions without compromising risk control.
The team started by analysing historical fraud cases and the current manual review process to understand decision bottlenecks, key risk indicators, and data availability. Building on this, they designed an LLM-based agent that triangulates structured internal data with unstructured external information, producing consistent and explainable risk evaluations within seconds. They iterated on this system together with Two’s team, focusing on validation, calibration, and integration into live workflows to ensure it reliably blocks fraudulent activity while allowing legitimate transactions to flow seamlessly.