Team 48

Clawback

Clawback is an AI agent that gets your money back after you buy, but only inside a Pact you signed: the AI proposes, a fixed rule gate decides, and every decision is signed and cites the store's own policy.

Check your own project

Community votes recognize the build and do not affect official judging.

Review the project

Start with the source code, then open the demo or video if available.

View GitHub repository
Project description
You buy something, the price drops a week later, and nobody claims the difference. Clawback does it for you, safely. You sign a Pact: save at least $10, ask me first above $150, never more than $250 at once. An open-weight model (NVIDIA Nemotron) reads receipts and proposes actions. It never decides. A deterministic gate (rules R0-R13) checks every proposal against the signed Pact (HMAC-SHA256). Change a rule and the signature breaks, so the next action stops. Approvals bind to a proposal hash, and a run-once guard means a repeated tap can't pay out twice. In the demo: $34 recovered on earbuds, a cheaper backpack refused (wrong colour), and a $199 lamp sent to "Needs your OK". Open "Why?" on any card to see the store's own policy sentence, retrieved through Senso, and the rule that decided. Senso: policy quotes with source IDs. No source found means the gate refuses to act. SingleStore: system of record, signed Pacts, audit log, run-once guard. Amplitude: every gate decision is an event, with stop reasons by rule and dollars recovered. Airbyte: Airbyte Core syncs receipts from a File source into Postgres. Persona: sandbox identity verification completed for the demo user; approvals re-check it before any money-moving action. NVIDIA open-weight model: proposals only. Scope: the store, support rep, receipts and Gmail link are simulated. The model proposes. The gate decides. Every action leaves an audit trail.
Tools used