Kunji

Why it exists

An AI that drafts replies is only worth having if you can trust it and measure it. Agents will never sit down and write test sets. But an agent who edits a draft has already shown what a better reply looks like. Kunji keeps that edit, and once a person approves it, it becomes a test case.

What I decided

I put the rules that must hold in code, not in the AI’s instructions, because an instruction can be talked out of firing and code cannot.

Fixed rules, written in code, block three kinds of mistake before a reply can go out, such as showing a customer’s personal data. Only those three block, because each one exposes something that cannot be taken back. Everything else only warns: agents stop reading a guard that blocks too often.

Kunji never sends anything by itself. Sending is always the agent’s own tap, and the server runs the three rules again at that moment, even on text the agent wrote.

Names and numbers become placeholders before any text leaves Kunji, and go back in on the agent’s screen.

The workspace scores replies with the same rules that guard them, so what Kunji measures and what it enforces cannot drift apart. Every score says who gives it: code, an AI judge or a person. No score rests on an AI judge alone.

My part

Kunji was my idea. I decided how it works and what it must never do, and I directed the build. An outside reviewer checked its test cases and the wording of its rules. It is a prototype, and it has never handled a real conversation.