SaveFlow

An AI agent that saves cancelling PayPal subscribers: it reads their reason and PayPal history, offers a pause, discount or cheaper plan within the merchant's rules, and executes it in PayPal.

Comment

Inspiration

Small businesses that bill monthly through PayPal Subscriptions (yoga studios, course creators, newsletter writers) lose subscribers they could have kept. Other payment platforms offer a retention step when someone clicks Cancel; PayPal Subscriptions doesn't, and popular retention tools don't support PayPal. When a PayPal subscriber cancels, they're simply gone. I wanted to build the missing save-the-subscriber layer for PayPal, using AI where it genuinely helps (understanding messy, free-text reasons in any language) while keeping it strictly inside the merchant's rules.

What it does

Subscriber clicks Cancel and explains why in their own words, in any language. The agent reads that reason plus their real PayPal history (tenure, payments, failed payments, current plan) and picks one offer: a pause, a discount or a cheaper plan, and writes a short, personal message. The offer is executed in PayPal. Pause = suspend with a scheduled activate; discount or cheaper plan = revise with the buyer's approval in a PayPal popup; otherwise a clean cancel. The AI never decides alone. Code computes which offers the merchant's policy allows; Gemini picks only from that list using structured output; code rejects anything else and falls back to a safe default, and rewrites any message that promises a different amount. Every decision is audit-logged with the exact PayPal calls, idempotency keys and PayPal debug IDs. Plain-English policy. Merchants write rules like "never more than 20% off, students get the cheaper plan, no offers for app problems". The AI turns them into structured rules, code validates them, and the merchant approves a summary written by code, not by the model. Rules it can't express (e.g. "give Jane a free month") are flagged as not enforced instead of being guessed. What-if simulator. Before a policy goes live, it replays past cancellations in an AG Grid and shows exactly which offers would change. Win-back. If someone cancels inside PayPal instead, the verified BILLING.SUBSCRIPTION.CANCELLED webhook triggers one personalised come-back offer; duplicate deliveries are ignored. Merchant workspace on AG Grid Community: every attempt with custom cell renderers, saved views, quick search, saved column layouts, CSV export, inline policy editing, and a drawer that shows the agent's work for any row.

How we built it

Backend: Python 3.11, FastAPI, SQLAlchemy, SQLite (re-seeded with labelled demo data on start). AI: Google Gemini with Pydantic structured output. Free-tier quotas are small (I measured 5 requests/minute and 20/day on one model), so the agent walks a chain of Gemini Flash-Lite models, each with its own quota, and caches the scripted demo answers. If every model is unavailable it falls back to a clearly labelled rule-based default, and the guardrail still applies. PayPal (sandbox): JS SDK Subscribe button; Catalog Products and Billing Plans (found-or-created by name, so restarts never duplicate them); Subscriptions API (show, transactions, suspend, activate, revise, cancel); Webhooks with signature verification. Every call carries an idempotency key and a timeout, and errors keep PayPal's debug_id. Frontend: React, Vite, TypeScript, Tailwind CSS, AG Grid Community. Engineering: an explicit subscription state machine (ACTIVE / PAUSE_SCHEDULED / SUSPENDED / CANCELLED) with valid transitions only; PayPal is the source of truth, so webhooks reconcile by re-reading the subscription (handles out-of-order events); approval confirmation is a conditional update, so a popup callback and polling that race apply a change exactly once. 36 automated tests. Hosting: one Docker service on Render's free tier serving the API and the built frontend.

Challenges we ran into

India-region PayPal sandbox accounts can't approve subscriptions (PayPal shows a generic "things don't appear to be working" page), so everything runs on US sandbox accounts. The README tells judges this. PayPal rejects a per-subscriber price override on the same plan (OVERRIDES_ON_SAME_PLAN_NOT_ALLOWED, found in my day-1 sandbox test). I model each discount as its own plan with a discounted trial phase that returns to the regular price automatically. Plan changes need the subscriber's approval in PayPal. I open approval in a popup and only mark the offer as saved after re-reading the subscription from PayPal. Verified live: after approval the subscription's plan changes immediately, and its next billing date stays the same. Keeping an LLM inside hard business limits: moving the limits out of the prompt and into code was the key design decision. Free-tier rate limits: solved with the model chain, cooldowns and a demo cache.

Accomplishments that we're proud of

0 policy violations in a 50-case eval (all 50 answered by Gemini), including 10 prompt-injection attacks: 10/10 flagged, 0 false flags, 92% reason accuracy across English, Hinglish, Spanish and sarcasm, and 50/50 offers that fit the reason. The model never proposed an out-of-policy offer in that run; the code guardrail that blocks one is tested separately with a deliberately misbehaving model. Real PayPal actions, verified end to end on the hosted demo: a live sandbox subscription paused by the agent (PayPal shows it SUSPENDED), and a plan switch approved in the PayPal popup and confirmed from PayPal. An honest dashboard: seeded demo data is labelled and tuned to the real-world ~15% cancel-flow save rate instead of an inflated number.

What we learned

AI agents are most useful when they choose inside rules a business already trusts. PayPal's Subscriptions API is already powerful enough to build retention on top of; the missing piece is the decision layer, plus the audit trail that lets a merchant see exactly why each offer was made.

What's next for SaveFlow

Try it with real merchants' cancellation history (with consent) to measure actual save rates. Learn which allowed offer works best per reason (a bandit over the allowed set; the statistics decide, the LLM explains). Failed-payment recovery: smarter retry timing and a friendly "update your payment method" flow. Verify the discounted trial phase on revise in the sandbox (currently untested and stated as a known limit). Try it Live demo: https://saveflow-ufbx.onrender.com (free hosting: the first load can take up to a minute to wake up). Sandbox buyer login for the real PayPal path is in the README. Code (MIT): https://github.com/Shiva-Preetham/SaveFlow-PayPal-AI-Hackathon Everything runs in the PayPal sandbox; no real money moves. "Sunrise Yoga Studio" is a made-up merchant.

Built With

Share this project:

Updates

Submission history