Kalinaw Bank
Live · Mesh API

Request inspector

Send a request as one of the bank's apps and watch the gateway handle it: classify the data, mask personal data, screen for injection, decide in code, call an approved model through Mesh, screen the answer and seal the audit entry.

Sample requests

Customer-facing·approved up toRestricted· risk high

The app asks; the policy decides which model actually answers.

297/2000

Patterns already see 1 email, 1 card number, 1 mobile number. Names and IDs are found by the model.

Up to four Mesh calls per request · policy v1

Gateway timeline

Pick a sample or write a request, then send it through the gateway.

  1. 1
    Identify app and callerCode

    Which registered app is calling, who owns it, and the highest data class it is approved for.

  2. 2
    Detect and mask personal dataCode gpt-oss-120b

    Patterns in code catch card, account, TIN, mobile and email numbers. An open-weight model finds names, addresses and IDs. What it finds sets the data class.

  3. 3
    Screen for injectionCode gpt-oss-safeguard-20b

    A safety classifier and fixed patterns score prompt injection and jailbreaks, and flag requests that decide something for a customer.

  4. 4
    Policy decisionCode

    Code applies the policy in a fixed order: block, reroute to an approved model, hold for human review, or allow. One rule decides.

  5. 5
    Model call through Mesh approved chain

    Mesh calls the chain approved for the data class and falls back only inside it.

  6. 6
    Screen the answerCode jev-1.13

    A second model scores disclosure and conduct; patterns re-check for leaked numbers; placeholders are restored for the app.

  7. 7
    Seal the audit entryCode

    One entry, hash-chained to the last: masked prompt, digest, class, decision, rule, model, tokens, cost.