Team 16

Owed

A personal agent that gets your money back after you shop: it reads your receipts, finds price drops, duplicate orders and double charges, files the claim with the store on its own, verifies the refund, and shows the exact policy sentence behind every decision.

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
Everyone builds agents for the two minutes of buying. Owed works after you buy. It reads the receipts in your Gmail (read-only), finds money you are owed, pulls the store's own policy, decides in code whether you are entitled, files the claim, and counts a case as done only when the store returns a refund ID. Every step lands in an audit trail. Demo run: 12 emails, 5 possible refunds, $263.50 recovered across 3 stores, 2 cases declined with the policy sentence that says why. One click, no human in the loop. Stores and policies are fictional demo data, labelled as simulated in the audit trail. Design: models read, code decides. No claim without a policy sentence found word for word. An edge-case inbox with 17 tricky emails decides 10/10 correctly. Tools: Senso holds the store policies, every decision cites a Senso passage and Senso fact-checks the agent's reading before it acts. Superlinked extracts receipts (GLiNER) and reranks policy passages (Qwen3). Statsig sets the autonomy limit ($150 per claim) and approval gate. Amplitude receives every step as an agent event. Cognee remembers past claims per store. SingleStore stores cases and the full audit trail, queryable in SQL. Live: https://owed-3tov.onrender.com - Code: https://github.com/Mertdem21/owed
Tools used