Switchproof

Bank payment migrations fail in public and expensive ways: TSB's 2018 migration cost about £81m in fines and compensation. SwitchProof is an AI test team for these migrations. A tester writes the payment rules in plain English. A coordinator agent on Vultr Serverless Inference (GLM-5.3) plans the work, and a generator agent (DeepSeek-v4.1-flash) writes ISO 8583 test cases. Nothing runs until the human approves every test. Approved tests, plus a 10,000-transaction replay of IBM's public synthetic TabFormer dataset, run in throwaway gVisor sandboxes on Vultr Kubernetes, in two isolated pools on separate VMs: agent-written tests and replay data never share a machine. When the new system misbehaves (in our demo a seeded bug charges customers twice), a triage agent designs its own follow-up tests to find the exact boundary, explains the root cause and the money at risk, and the human blocks or approves the release. A real run on Vultr: 10,022 tests in 42 sandboxes, 0 errors, defect found in 8.9 seconds, $0.04 of AI cost. An isolation probe inside a real sandbox shows rm -rf /, internet, cloud metadata and escape attempts all blocked. The payment switches are simulated and the defect is seeded; everything else runs live on Vultr.

Note: the demo URL is substituted with the GitHub repository link.