Countless

Friends and family pay a disabled person's bills straight to the vendor through PayPal, so SSI never counts it as income. AI reads the bill; an agent verifies every payment.

Comment

Inspiration

SSI pays at most $994 a month in 2026, and Social Security counts cash gifts as income. So if a brother sends his sister $500 to fix her car, her check shrinks by over $400 two months later. But if he pays the mechanic directly, she loses nothing. Almost no family knows that rule, and there's no simple way for a few relatives to chip in and pay a shop together. We wanted to make the right choice the easy one.

What it does

Countless is a small circle of family and friends around one person on SSI (our demo is Rosa).

  • Snap the bill. A Groq vision model reads the shop, line items, total and the shop's PayPal email in about 3 seconds. Code checks the line items add up to the cent. The AI only reads; it never decides.
  • See the real cost. A rules engine built from Social Security's own manual (POMS) shows Rosa's next three checks for each choice. Cash: her January check drops from $1,029 to $607.27. Paid to the shop: $0 lost. Every number links to its rule.
  • Split it fairly. The bill is split across the circle, to the cent, within each person's monthly limit.
  • Save PayPal once with PayPal Vault. No checkout every time.
  • A person approves. One tap charges each saved PayPal its share (Orders v2) and pays the shop the full amount with PayPal Payouts in the same run. Rosa never touches cash.
  • An AI agent double-checks. A read-only agent uses the PayPal Agent Toolkit to look up every order and the payout in PayPal itself, then writes a plain receipt for the family. It has no tool that can move money.
  • One ledger built with AG Grid shows every dollar in and out. Countless always holds $0.

How we built it

Next.js 16, TypeScript and Tailwind. PayPal REST APIs in sandbox (Vault v3, Orders v2, Payouts) plus the PayPal Agent Toolkit. Groq models through the Vercel AI SDK: qwen vision for reading bills and gpt-oss-120b for the audit agent. PGlite locally and Postgres in production, and AG Grid for the ledger. The SSI engine is plain, tested code (26 tests in total).

Challenges we ran into

Our first design charged each helper with the shop as the payee, so the money would never touch the platform. PayPal returned 422 BILLING_AGREEMENT_NOT_FOUND, because a saved PayPal belongs to the platform that saved it. We redesigned it so shares are captured and paid out to the shop in the same run, and the ledger proves the balance goes back to zero.

The SSI rules were the other hard part: the two-month lag, the $20 and $60 exclusions, and the 2024 change that removed food from in-kind support. We checked every number against SSA pages.

Accomplishments that we're proud of

  • A real end-to-end sandbox run: 4 Vault charges plus a Payouts transfer to the shop, from one Approve tap.
  • The agent verified all 5 payments with PayPal in about 10 seconds.
  • Every dollar figure on screen has a citation behind it.
  • An accessible design: high contrast, large tap targets, and plain language for older relatives.

What we learned

AI belongs in money tools where it reads messy documents and checks work, not where it makes rules decisions. We also learned a lot about how PayPal handles multi-party payments and vaulted accounts.

What's next for Countless

Letting anyone start a new circle, warnings for rent bills (shelter still counts as in-kind support), a Spanish version, and a real ABLE account connection.

Sandbox only. Not legal or benefits advice.

Built With

Share this project:

Updates

Submission history