Coinbase details AI fraud agent architecture for Onramp

Coinbase published a three-part engineering series on the fraud agent guarding Onramp: a fine-tuned Qwen3.5-9B beat Opus 4.5 on precision, recall, F1 and dollar-weighted recall by 7.5–35.4 points. In an online A/B test the agent cut fraudulent transactions by 30% and fraud value by 22%. Training cost under $100 on a single RTX PRO 6000 GPU.

• Fine-tuned Qwen3.5-9B beat Opus 4.5 on four metrics by 7.5–35.4 points
• A/B test: 30% fewer fraudulent transactions, 22% less fraud value
• Median latency 0.683s vs 1.515s for Opus 4.5
• Training cost under $100 on one NVIDIA RTX PRO 6000 96GB GPU