Loops HouseHomeroll

NEAR HACKATHON
Links
Team
1 member- ZEOwner
Ze Quan
Overview
Problem: Malaysian SMEs wait about 64 days for invoices to be paid, but wages are due every month. Their foreign workers then lose 7.51% on average sending money home to Bangladesh. AI agents could handle the paperwork, but a fooled agent must never be able to move money.
What we built: Homeroll, a web app with four Rust contracts on NEAR testnet. An SME owner checks an invoice with an AI agent and takes an advance from a lender pool, capped by the invoice’s grade. The owner runs payroll: the payroll contract checks all 40 pay lines (active payees, wage caps, once per period, treasury buffer) before anything is paid. Workers see their wage on their phone and send money home for about 1.1% all-in, with no seed phrase and no gas. Lenders fund the pool from any chain.
How NEAR fits in: the AI agents hold function-call access keys. The payroll agent can call only submit_run and pay_run and cannot attach a deposit, so when it tries to call the token contract NEAR rejects the transaction before it is included. Our /lab runs six attacks that fool the agents (a stranger’s wallet, fake overtime, double financing, hidden prompt text, a direct token call, a replayed payout) and none moves money. Workers sign meta-transactions (NEP-366) so a relayer pays their gas. The owner approves payouts with NEP-413 signed messages. Worker and lender accounts are random pseudonymous sub-accounts, so no names reach the chain. NEAR Intents gives live quotes for Send home and for lender deposits from other chains.
Mocked and labelled in the app: the invoice model, banks, MyInvois, DuitNow, bKash and the Confidential Intents layer. All data is synthetic and no real money moves.