Kusubi

What Kusubi does, step by step.

Your applications, assistants and agents send their AI requests to one single address. Before a model answers, every request goes through a series of checks. You stay in control, and you keep the proof.

Six steps, for every request

  1. 1

    Check who is asking

    Does the request come from a known user, with a valid key? Is its behaviour normal?

  2. 2

    Understand the request

    Kusubi recognises what it contains: personal data, sensitive information, type of task. It adds labels to it.

  3. 3

    Apply your rules

    Your rules apply to these labels. The request is accepted, adapted or refused. What must not leave does not leave.

  4. 4

    Protect personal data

    Names, addresses, numbers: they are replaced before being sent to the model, then put back in place in the response.

  5. 5

    Choose the model

    The request goes to the cheapest authorised model that meets the expected quality.

  6. 6

    Measure and keep the proof

    Every decision is recorded: who, what, which model, why, at what cost.

Your administrator stays in control

From an administration area, they manage users and their keys, write the company's rules, and see every decision with its reason.

Four requests, four decisions

Fictitious examples. Some requests go through, others are refused: the reason is shown to the user and to their administrator.

  1. Request 1 : “Summarise the minutes of Monday's team meeting.”

    1. Who : passed
    2. Content : passed
    3. Rules : passed
    4. Data : passed
    5. Model : passed
    6. Proof : decision recorded

    ✓ Accepted

    What the user sees

    Here is the summary.

    What the administrator sees

    Accepted · text summary · no sensitive data · standard model

  2. Request 2 : “Draft a reply to Ms Martin about her contract no. 4521.”

    1. Who : passed
    2. Content : passed
    3. Rules : passed
    4. Data : data replaced
    5. Model : passed
    6. Proof : decision recorded

    ✓ Accepted

    What the user sees

    Here is the draft reply. The name and the contract number were not sent to the model.

    What the administrator sees

    Accepted · personal data replaced before sending, restored in the response

  3. Request 3 : “Analyse this file of the team's salaries and compare them.”

    1. Who : passed
    2. Content : passed
    3. Rules : Refused
    4. Data : not reached
    5. Model : not reached
    6. Proof : decision recorded

    ✕ Refused

    What the user sees

    Request refused: it contains payroll data. Your company's rule reserves this data for the HR department.

    What the administrator sees

    Refused · rule “Payroll data” · reason shown to the user

  4. Request 4 : “Here is our server access key: sk-…, check the configuration.”

    1. Who : passed
    2. Content : passed
    3. Rules : Refused
    4. Data : not reached
    5. Model : not reached
    6. Proof : decision recorded

    ✕ Refused

    What the user sees

    Request refused: it contains a secret (an access key). Remove it, then send the request again.

    What the administrator sees

    Refused · secret detected · leak prevented · alert to the administrator

The animation stops when the pointer or the focus is on it. With “reduce motion”, the four examples are shown one after the other.

Let's talk about your AI