Inspiration
Editors and motion designers do the work, then wait. 71% of freelancers have struggled to collect payment at least once (Freelancers Union). The brand goes quiet, "make it pop" revisions never end, and the subcontractor gets paid last, if at all. Underneath all of it is one problem: nobody wrote down a payday the code can keep.
What it does
yv. is a content-ops workspace for brands and the creative crews they pay. In the Approved = Paid flow:
- Both sides sign the deal: milestones, amounts, each person's share, the included revision rounds and a review window. The terms are hash-locked.
- Each milestone is funded with a PayPal hold (Orders v2 AUTHORIZE), so the money is set aside but not charged.
- When a cut is uploaded, automatic AI checks run, powered by an NVIDIA NIM language model and FFmpeg. FFmpeg measures length, loudness and captions. The model explains any failure in plain words.
- The reviewer pins comments to exact timestamps. Two rounds are included, and an extra taste round becomes a priced change order sent with PayPal Invoicing.
- The Fair Review Clock counts business days (skipping the brand country's bank holidays) and sends reminders at 50% and 80%, including to a backup reviewer. If nobody answers, the deal's own rules apply: a grace period, then release. A real dispute pauses the clock.
- On approval, the hold is captured (Payments v2), and PayPal Payouts pays every crew member their signed share in one batch.
- AG Grid powers the brand's Payables board and the crew's "Where's my money" view.
- Brand Brain, the team's AI assistant, remembers the brand and drafts briefs. A deal copilot answers money questions but never moves money.
AI checks facts. People own taste. The deal decides the money.
How we built it
- Web: React + Vite on Vercel. API: Node + Express on Render, the only writer to MongoDB Atlas. Agent: Python FastAPI + LiteLLM on Render (River AI Qwen, NVIDIA NIM fallback).
- PayPal: Orders v2, Payments v2 capture and void, Invoicing v2 and Payouts, all in the sandbox. Money calls go through a typed PayPal Server SDK generated with APIMatic, with idempotent request ids so nothing is charged twice.
- Every money move is first dry-run against policy rules P1 to P9 (state, actor, exact amount, locked terms, approved version, human tap, currency, live authorization). Blocked attempts are written to a hash-chained ledger as "No money moved".
- The review clock runs on Render, inside the API's Money Agent tick.
Challenges we ran into
- Our first PayPal sandbox app was India-based, and Payouts were refused (PAYOUT_NOT_AVAILABLE). We moved to a US Business sandbox app to make real sandbox Payouts batches.
- Judges shouldn't need a PayPal login, so the demo draws its holds from a pool of authorizations a sandbox buyer approved ahead of time. Capture and void on them are real sandbox calls with real ids.
- Keeping AI away from money: models name cards and ids, but every amount is computed in Node from the signed terms, and any agent request that carries an amount is refused.
Accomplishments that we're proud of
- A full loop on the PayPal sandbox with real ids: a hold, AI checks, review, approval, capture, then one Payouts batch split 70/30.
- A review clock that turns brand silence into a written rule instead of a lost payment.
- Over 800 API tests, including a policy red-team suite.
What we learned
The hard part of freelance pay isn't sending money. It's agreeing in advance when money moves, and having code that keeps that promise.
What's next for yv. Approved = Paid
Deploying the review clock as a Render Workflow (it's written but not deployed yet), live PayPal onboarding for real brands, and more checks such as brand colours and logo safe zones.
All payments in the demo run in PayPal Sandbox. No live money moves.
Built With
- ag-grid
- apimatic
- express.js
- fastapi
- ffmpeg
- litellm
- mongodb-atlas
- node.js
- nvidia-nim
- paypal
- paypal-orders
- paypal-payments
- paypal-payouts
- python
- qwen
- react
- render
- river-ai
- socket.io
- typescript
- vercel
- vite