CheckWithUs

CheckWithUs is scam protection where the older adult does nothing new and the family does the work. On the older adult's iPhone, an on-device scam model inside a Message Filter extension moves scam texts to Spam before they are read, a Call Directory extension labels known scam numbers "Possible scam", and every payment goes through fixed rules: a risky one pauses and the phone says "Your money is safe. Your family will call you." The family side — an iPhone app and a dashboard at checkwith.us — shows the amount, the payee and each hold reason in plain words, and approves with Face ID or denies with one tap.

The governing rule is that AI suggests and code decides money. One pure function, decideHold, pauses a payment for reasons such as a gift-card, crypto or wire-transfer category, cash at an ATM followed by a gift card, a payee never paid before, an amount over the family's limit, spending far above usual, or payment during a phone call — and the on-device Guard score can only add a hold reason, never remove one. Over $1,000, two family members must say yes and one no stops the payment; one tap freezes the whole spending balance into a vault account; scoped allowances let routine bills (for example utilities up to $400 until 30 November) pay without a push. Money is integer cents everywhere, and the older adult's texts and bills never go to a cloud AI model.

The scam model itself is deliberately small enough to run where the texts are: Guard tiny, character and word n-grams with logistic regression, weighs 1 MB and scored 96.0% scam recall with 2.4% false holds on 884 real test texts. Underneath, Capital One Nessie is the sandbox bank (accounts, vault, merchants, purchases, bills, withdrawals, transfers), Auth0 handles login, roles and the bill-pay agent's single pay:bills scope, and Tiger Data holds the spending ledger whose continuous aggregates power the unusual-spending rule. The live site's "Try a payment" widget runs the same decideHold code in the browser.

Vibe-coded evidence, in the authors' own words on DEV (18 October 2026): "We built with Claude Code and OpenAI Codex CLI. Claude Code wrote most of the code." The team of four built it in 36 hours at SwampHacks XII (16–18 October 2026), and the code is public. Honest limits, as the authors state them: this is a hackathon build — the demo's Sam is a made-up older adult, the bank is a sandbox rather than a real bank, there is no server push yet (alerts arrive by polling while the app is open or recently used), no usability test with older adults has been run yet, and the reported scores describe their test set, not every scam. GitHub stars: 0 at the 28 October 2026 check.