Inspiration
Almost everyone is paying for something they've forgotten about: a streaming plan that quietly went up in price, a gym pass nobody uses, two apps doing the same job. The money leaves in small amounts, so nobody notices. We wanted an AI agent that actually does something about it, but one people could trust, because "an AI that can spend your money" is a scary sentence.
What it does
Autopay Bouncer reviews your recurring payments and flags four kinds of problems: price hikes, subscriptions you've stopped using, overlapping services, and failed payments. An AI model explains each finding in plain English and suggests an action. Nothing happens until you click approve. When you do, the app suspends the subscription through PayPal, and you can undo it with one click. Every action goes into a log.
On our demo data it found 57 dollars a month of waste:
$$57 \times 12 = 684 \text{ dollars per year}$$
How we built it
- Backend: Python and FastAPI
- Detection: plain, testable code that checks transaction history for price changes, duplicates by category, failed payments, and long gaps in use
- AI: Chatgpt 4-o mini writes the explanation for each finding
- PayPal: the Subscriptions API in the sandbox (OAuth, suspend, reactivate and cancel)
- Frontend: a single plain HTML and JavaScript page
The key design choice is the trust boundary. The AI has no tools and no access to PayPal. It can only write text. The only code path that calls PayPal's suspend or cancel endpoints is the approval endpoint, which a human click triggers. Even if the model misbehaved, it couldn't move anything.
Challenges we ran into
- PayPal's Subscriptions API is built for merchants. It lets a business manage the subscriptions it sells, not the ones a person buys from other companies, so a consumer app can't pull your subscriptions to other services through it. Our demo therefore mixes seeded data for the full story with two real sandbox subscriptions that the agent can genuinely suspend and reactivate. We say so plainly rather than faking it.
- Sandbox setup. The subscription approval page kept failing with a generic error until we recreated our sandbox accounts with a US country setting.
- Keeping the AI in its lane. We moved all detection into deterministic code so the model only explains, which made the results testable and the behavior predictable.
Accomplishments that we're proud of
It works end to end. The agent flagged four problems, and approving one really changed the subscription's status in PayPal's sandbox dashboard. The AI recommends, and the human decides.
What we learned
How PayPal's subscription lifecycle works, how to design an agent around a hard human-approval boundary, and that the most useful thing an AI agent can do is often to explain a decision clearly rather than to make it.
What's next
- Importing a real PayPal activity export (CSV) so anyone can audit their own recurring charges
- A weekly check-in that asks whether you still use each service
- Webhooks so the app updates itself when PayPal reports a change
- More detectors, such as free trials about to convert
Built With
- css3
- fastapi
- html5
- javascript
- oauth2
- openai
- paypal
- paypal-subscriptions-api
- python
- rest-api
- uvicorn
