Don't trust us. Check.
Kavach promises the AI can never pay out on its own. This page lets you test that promise yourself, live, against the real database. Nothing here is typed in by hand.
- 1Press Run on Proof 1Postgres refuses the AI's attempt to write a refund.
- 2Press Run on Proof 2Edit one row of a copy; the check finds that row.
- 3Open any receipt belowEvery check behind a real decision.
Free to read and reason. Never free to pay.
- Understands the message (LLM, 3 languages)→
- Decides what to look up→
- Reads the records→
- Proposes an action→
- Verifier re-checks→
- Allowed · seller · blocked
- • choose which records to read (order, courier, policy, photos, refunds)
- • handle two requests in one message
- • ask a numbered question when something is missing
- • remember the chat, even across restarts
- • write to the money tables (Postgres refuses: Proof 1)
- • skip a check: a safety rule adds any record it forgot
- • invent facts: replies are filled in from the records
- • act on “the owner said so” or instructions hidden in data
The AI's process cannot pay anyone
The chat and AI run as a database user that Postgres won't let write refunds. Press Run and we try anyway.
Why it matters: Even if someone tricks the AI completely, it has no way to move money.
The decision log can't be changed quietly
Every decision is saved with a fingerprint (SHA-256) that includes the row before it. Press Run: we copy the log, change one word in a random row, and re-check.
Why it matters: Nobody, including us, can quietly rewrite what Kavach decided.
Every payout needs a one-time signed slip
The verifier signs the exact order and amount (HMAC). The slip works once and expires in 10 minutes. Open a receipt to see a real one.
Why it matters: An approval can't be replayed, or reused for a bigger amount.
Every decision, live
web · WhatsApp · Telegram- No decisions yet.