Browser-based transaction analysis
A mock banking form scored by a small ONNX model (logistic regression: MatMul → Add → Sigmoid over 7 normalised features) running in Crucible's Rust WebAssembly runtime. The model file is fetched once; scoring happens in this tab and sends nothing over the network. The model was trained on synthetic transactions (see models/fraud/README.md), so its scores illustrate the runtime, not real fraud risk.
Transaction input
Pick a scenario or enter values by hand.
No inference yet
Fill the fields and press Analyse transaction. The ONNX model executes inside this tab.
Why on-device inference matters for finance
Under PCI-DSS and GDPR, raw cardholder and account data must be tightly scoped. Sending features to a hosted inference gateway expands the compliance boundary: the model host becomes an in-scope service provider, and cross-border transfer obligations follow.
This demo shows the alternative topology: the model ships as a static file, is evaluated in the browser's WebAssembly sandbox, and the app only needs the resulting score. Scoring involves no network round trip.