Team 66

Glassbox

Glassbox is a personal shopping agent for people who won't hand an AI their card blind: it compares offers using your memory and rules, rejects fake reviews and bait prices, waits for your signed approval, and logs every step to a tamper-proof audit trail.

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
For anyone who wants an agent to shop for them but can't trust a black box with their card. Give Glassbox a task like "noise-cancelling headphones for long flights under $300". It runs a 9-step loop: parse intent, recall your memory (it knows you wear glasses and that you returned narrow Kestrel shoes), load your guardrails, pull offers from 5 merchants, and verify each one: ratings discounted for fake reviews (a 4.8★ product with 46% fake reviews is rejected), price-spike detection against 30-day history, merchant trust, and delivery deadline. It then ranks options on a 100-point score where every point is explained. It never spends silently. It drafts a signed spending mandate (item, merchant, max amount, 15-min expiry) and waits for you. Persona verifies a human approved it, the live price is re-checked to stop bait-and-switch, and the order is placed. Every step, with its evidence, is appended to a SHA-256 hash-chained ledger. Verify it in-browser with WebCrypto; one click simulates tampering and shows the exact broken event. It passes 10/10 adversarial safety drills. Partners: Senso (context), Airbyte (merchant data), cognee (memory), Superlinked SIE (ranking), Arcee (reasoning), Statsig (guardrails), Persona (human verification), SingleStore (ledger store), Amplitude (analytics), Firedrill and Guild (drills and evals).
Tools used