Inspiration
Gig workers wait days to get paid, and marketplaces lose time reconciling earnings and chasing payment exceptions by hand. We wanted AI to do the checking and matching while strict rules, not the AI, decide who gets paid.
What it does
GigFlow AI takes a marketplace earning and checks it against a payout policy. It then pays the worker through PayPal Sandbox, reconciles the payment, and routes any exception to a human, with an audit trail at every step. A payout is approved only if 0.01 ≤ amount≤100,0000.01≤amount≤100,000, the currency is USD, and the earning is in an allowed state.
How we built it
A FastAPI backend stores everything in SQLite. An OpenAI function-calling agent gathers context through domain tools and the PayPal MCP server, and the React + Vite dashboard shows the results. Payouts always go through a deterministic policy gate and the backend's PayPal service, never through the model.
Challenges we ran into
Preventing double payouts under concurrent requests took optimistic-concurrency updates and idempotent ingest. Making sure the AI could never authorize money movement, and running the PayPal MCP server as a Node subprocess, were the other hard parts.
Accomplishments that we're proud of
One POST /api/earnings call ingests, settles and reconciles an earning end to end, with a full audit trail. The 84-test backend suite passes without calling live PayPal or OpenAI.
What we learned
AI works best as an investigator and router, with hard rules guarding the money. Design for retries and races from day one, because payment APIs fail in ways you only see under load.
What's next for GigPay
Multi-currency payouts, live PayPal webhooks, authentication with roles, and a production-grade database. We also want to add dispute and refund automation.