Graph Transformer based Anti-Partner Fraud Guard

**The Problem** Traditional anti-fraud systems treat every partner as an isolated row. They check thresholds — "Did commission exceed X?" — but affiliate fraud rings operate as coordinated networks: fake clients, mirrored trades, timed deposits. Viewed row-by-row, everything looks normal. In our analysis, 60% of all commissions ($61M out of $101M) flowed to fraudulent partners that rule-based systems would miss. **Our Solution: Deriv Guardian** We shift fraud detection from row-level to graph-level with three layers: **1. The Brain (Kumo.ai Relational Graph Transformer):** We model the affiliate ecosystem as a graph — 2,358 account nodes connected by 6,883 timestamped edges (referrals, trades, commissions). The model performs message passing across the network, learning relationship patterns — not just transaction features. Result: 94% precision, detecting 65 out of 68 fraudulent partners . **2. Temporal Intelligence:** A timeline visualization enables "time travel" through trading data, revealing how fraud rings evolve from grooming (small test trades) to active fraud. The graph transformer detects structural anomalies 2-3 weeks before traditional systems. **3. GenAI Copilot (Azure OpenAI):** Click any flagged partner and AI generates a 3-bullet investigation report — Evidence, Impact, Recommendation — reducing investigation time from hours to 30 seconds. **Data Engineering:** No public affiliate fraud dataset exists, so we built one. From IBM AMLSim (5M transactions, 518K accounts, 370 fraud rings), our Python pipeline assigns Partner/Client roles via network in-degree, maps to Deriv's schema with 7 instruments, and injects realistic fraud: 714 opposite trades (mirrored BUY/SELL within 60s) and 221 coordinated bonus abuse deposits. **Tech:** Kumo.ai (Graph Transformer), Azure OpenAI GPT-4o, FastAPI, React, Python/Pandas,Docker **Impact:** $61M in preventable fraud identified. Early detection during grooming phase. Investigation time: hours → 30 seconds.