X | Devpost

The card she signs: her own words, who, and how much. The AI drafted it, but it has no tool that moves money.
"Sent, verified" appears only after PayPal confirms the payout, by a signature-checked webhook or a status check.
Her daughter Anna is the family guardian: she approves a first payment to someone new, or freezes everything at once.
How Star checked this: every PayPal API call it made, and each safety check the payment passed.
What Star stopped: every blocked request, the rule that stopped it, and zero PayPal calls for all of them.
A "safe account" scam script is stopped before any model sees it: the fraud hotline, and 0 money-moving PayPal calls.
What Star stopped: every blocked request, the rule that stopped it, and zero PayPal calls for all of them.

Inspiration

For many older adults, paying someone is harder than it should be. Payment apps mean small text, menus, passwords, one-time codes and typing an email address exactly right. So they wait for a grandchild to help, or hand their phone or card details to someone else, which is exactly where mistakes and scams get in.

But talking is something everyone can do. We wanted Mrs. Lee to say "Send 50 to my daughter" and have it done, as easily as asking a family member, with Star answering her out loud in plain words.

Making payments easy for older people also makes them easier to abuse. An assistant that can pay can mishear an amount, and it can be talked into paying by a scammer on the phone. So Star is built ease-first, with a safety net that stays out of her way until it is needed.

What it does

Star is a voice payment assistant for older adults and their families. It does one job well: it sends money to her family when she asks in her own words, through PayPal Payouts.

She just says it. Mrs. Lee (fictional, 74) presses Speak and says "Send 50 to my daughter". No menus, no forms, no email addresses: Star knows that "my daughter" is Anna Lee.
Star says it back. Star replies out loud in short, plain sentences and shows one clear card: "fifty dollars ($50.00) to your daughter, Anna Lee". Care mode, the first thing Star offers, uses larger text and simpler words.
One touch to confirm. She confirms with her fingerprint or face, a passkey on her own device. No password, no code to copy.
PayPal sends it, and Star tells her when it arrived. "Done" appears only after PayPal itself confirms the payment, by a signature-checked webhook or a status check.

Her family is part of it. Her daughter Anna is her family guardian. On her own phone she approves a first payment to someone new, refuses one, or freezes everything at once. Lifting a freeze needs both of them.

Protection is built in, not bolted on. The model understands her words but has no tool that moves money: it can only draft a card. Fixed rules check the draft against what she actually said (the person, the amount, someone new, limits, the time of night), so a misheard or wrong amount is frozen before she ever sees it. Classic scam scripts, such as "The police say I must move my savings to a safe account", are stopped before any model sees them, with the fraud hotline shown and 0 money-moving PayPal calls.

The Dashboard, built with AG Studio, is the family's view: what she sent and to whom, whether PayPal confirmed each payment, and what Star stopped and why. A custom widget, Safety path, shows every step a request takes with live counts. The built-in AI assistant answers questions about it and builds widgets ("Who got the most?").

How we built it

Voice: speech recognition turns her words into text, and every reply is also read aloud, with amounts said in words. Care mode changes the size and wording on screen, never the safety checks.
PayPal: Payouts (create, batch and item status, cancel unclaimed items), signature-verified webhooks (/v1/notifications/verify-webhook-signature), OAuth client credentials, idempotent request ids, and reconciliation of unknown outcomes. Orders, captures, refunds and Invoicing are implemented and tested too, but kept off the screen so the demo does one thing well. A Postman collection replays the transfer path against the sandbox and shows Star refusing an unsigned card.
AI: DeepSeek V4.1 Flash, an open-weight model hosted by our university and called through its OpenAI-compatible Chat Completions API with function calling, understands her requests, drafts cards and answers questions. A second, independent reading of her words can freeze a draft, and a separate scam assessment can only make things stricter, never looser. Offline rules run the same pipeline without a model and say so on screen. The core rule: the model understands, code decides, she authorises.
Device signatures: Web Crypto keys, or real passkeys (Touch ID, Face ID, Windows Hello), sign the digest of the exact card she sees. Change one character and the signature no longer fits.
AG Studio: a React + TypeScript dashboard with three tables, a calendar, a currency format, a measure, a theme taken from the app's own colours, a custom widget, and the Studio Agent Framework running on Star's model through a Chat Completions adapter we wrote and a capped server proxy.
Stack: Python, Starlette, vanilla JavaScript, SQLite, AG Grid, AG Studio, Vite and Playwright, deployed on Render.

Challenges we ran into

Keeping it simple for her while keeping it safe. Every extra question or button is a barrier for an older user, so the checks run silently, and she only hears from them when something is actually wrong.
A payment request can succeed while its reply is lost. Treating that as a failure invites a second payment, so an unknown outcome stays visible, keeps its allowance reserved and is reconciled with the same request id.
AG Studio writes inline styles whose values change with the data, while Star's page allows none. We run the dashboard in a same-origin frame with its own policy, so the app keeps its strict Content Security Policy.
Our model speaks Chat Completions, while AG Studio's example adapter speaks the Responses API, so we wrote our own adapter.

Accomplishments that we're proud of

A real sandbox payout from a spoken sentence, said back to her in plain words, signed on her device and confirmed by PayPal, on the hosted demo.
A planted 10x model bug ($500 drafted for her $50) is frozen by code and never reaches a card she can sign.
701 unit tests, 150 end-to-end browser checks and 94 evaluation cases pass on the current code.

What we learned

For older users, convenience and safety are not opposites. The easiest experience is one where she can speak naturally, and the safest design is not a smarter model but a path where the model's mistakes have nowhere to go. We also learned that "processing" and "unknown" must never look like success, or like a failure that is safe to retry.

What's next for Star

Testing with older adults and their families, more languages and dialects, real notifications to the guardian, and an external security review before any live money.