The problem Tiba solves

When software needs to pay a person, a human still has to approve every transfer. That works at ten payments a day and breaks at ten thousand, and nobody gets paid at 2am or on a weekend. Meanwhile the person waiting is a freelancer or contractor who did the work, and the person on the hook if the software pays the wrong claim is the operator who switched it on.

Tiba lets software pay a person without a human approving each transfer, and keeps that safe with limits set once in advance. When an agent asks to pay someone, Tiba runs two separate checks through GonkaRouter: one reads the delivery note or invoice, the other reads our own records and never sees the note. Both have to name the same person and the same amount. If they disagree, Tiba refuses. No averaging, no tie-break. Then the operator's limits apply — a cap per payment, caps per hour and per day, a list of who can be paid, and a kill switch — and the payment settles on the Sui test network with a public receipt showing exactly why it was paid or refused.

Proof, on the test network, not a mock: 20 out of 20 legitimate payouts paid, 10 out of 10 tampered claims refused, no false refusals, about 13 seconds per decision.

Not included: real money, mainnet, KYC, cross-border compliance, image-based evidence. Settlement uses the network's own coin as a stand-in for USDC until test USDC is available.

AI tools used during development. This project was built by a team in which the founder directs AI coding agents rather than writing most code by hand. Tools used: Claude Code (Anthropic) — primary agent for architecture, implementation, review, and documentation; OpenAI Codex CLI — delegated implementation slices, research, and independent review of Claude Code's output. GonkaRouter is used inside the product as the verification layer, not as a development tool. Problem selection, product scope, every design decision, the pitch, and the demo were human-directed. No part of this submission is represented as entirely human-written code; most commits were agent-generated under human direction inside the hackathon window.

Challenges we ran into

The first working version took 45 seconds per payment — too slow for a stage demo. The two checks were running one after the other instead of at the same time; fixing that was the first win.
Response times on the router varied about tenfold depending on which node answered. We now send each check to two models at once and take the first valid answer.
One model kept "thinking out loud" before answering and ran out of space before giving us anything usable, so we dropped it.
Another returned oddly padded answers when we asked for strict JSON, so we validate every answer ourselves instead of trusting the format.
The router doesn't accept images, so evidence is text only for this submission.
Our deployment shipped with no styling for a while because a build plugin was missing. Nobody had looked at a rendered page — only at status codes. Lesson learned.