IntentPay

Problem

Wallets expose irreversible actions through low-level addresses, gas settings, and contract calls, which is difficult for non-technical users and risky for teams that need guardrails.

System

A React client and Express API resolve payment intent through an LLM adapter into structured transaction plans, with encrypted wallet data, spending guardrails, and the wallet kept as the final execution boundary.

Proof

• Separates AI planning from irreversible on-chain execution; the wallet remains the approval step.
• Handles named recipients, split payments, transaction previews, gas estimation, and transaction history.
• Uses AES-256-GCM encrypted wallet data, bcrypt password hashing, and security event logging.

Architecture

1. DescribeThe user states a payment intent in natural language.
2. ResolveThe API resolves intent, contacts, balances, and guardrails.
3. PreviewA structured transaction plan is shown for review.
4. ExecuteThe wallet remains the final execution boundary.

Decisions

• Never let the model silently execute transactions; plans require explicit wallet approval.
• Resolve contacts and handles server-side before building the transaction preview.
• Treat suspicious or failed actions as first-class security events.

Tradeoffs and limitations

• The review step adds friction in exchange for a clearer safety boundary.
• Prototype scope targets the Sepolia network and a limited set of transaction types.

Technologies

• React
• Express
• Ethers.js
• LLM adapter
• Sepolia
• AES-256-GCM