Realtime Transaction Fraud Detection - with an LLM?
Ubuntu, Wednesday, September 30th, 2026
Canonical builds real-time fraud scoring with an open-weight LLM classifier on its open source MLOps stack.
Canonical demonstrates real-time transaction fraud detection with an LLM used as a classifier rather than a text generator.
Following TypeSafe AI's Jev service, open-weight models such as system-one-qwen3.5-4b-scorer read context and return probability scores in a single pass, which suits checkout-time middleware.
The walkthrough uses Canonical's open source MLOps stack, including Kubeflow, Katib, Feast, MLflow and KServe, from fine-tuning through serving.