Escrow Agent

Inspiration

Most small deals (freelance gigs, custom orders, resale) are agreed in WhatsApp or Instagram chats, with no protection. The buyer fears paying first and getting nothing. The seller fears working first and never being paid. Traditional escrow is slow, manual and built for big transactions, so people just trust each other and sometimes get burned.

What it does

Escrow Agent turns the chat where you agreed a deal into a protected payment.

• Chat to contract. Paste the conversation. AI extracts the price, the payment stages and objective "done when" conditions for each stage, and warns about gaps that usually cause disputes (vague words like "modern", no limit on revisions, unclear fees).
• PayPal holds the money. The buyer approves a PayPal order created with intent=AUTHORIZE. The money is held, not captured, so the seller can see it is secured but cannot touch it.
• AI checks the work. For each stage the seller submits a description and proof (images, SVG, text). The AI inspects the files, compares them to the agreed conditions, and returns release, reject or review, with reasons and a description of what it saw.
• Money moves only on a pass. A pass triggers a PayPal partial capture for that stage. A reject voids the remaining authorization, so the buyer gets the rest back. If the AI is unsure (or fails), the buyer decides. Nothing is ever auto-released on an AI error.
• Share links, roles and an audit trail. The seller sends the buyer a link. Buyer and seller see different actions, and every step is logged.

How we built it

• Frontend and API: Next.js (App Router, TypeScript, Tailwind), with a custom design system.
• PayPal (sandbox): Orders v2 (create with AUTHORIZE, payer-action approval, authorize), Payments API (capture an authorization with a partial amount and final_capture on the last stage, void an authorization), OAuth2 client credentials.
• AI: Google Gemini (gemini-3.5-flash-lite) in JSON mode with image input, for both the chat-to-contract extraction and the vision-based verification. A provider-agnostic layer also supports Claude, and there is a rule-based fallback so the app still runs without an AI key.
• Safety: chat text, delivery text and any text inside images are treated as untrusted data (prompt-injection resistant). Uploads are type- and size-limited and served with a sandboxing CSP. Amounts are capped, and the return URL only trusts the order the server created.
• Testing: end-to-end runs with Playwright against the real PayPal sandbox, including the buyer login and approval.

Challenges we ran into

• Getting PayPal authorizations, multiple partial captures and a void to line up with per-stage state. A rejected middle stage has to void the remainder but keep what was already paid.
• Making the AI judge evidence, not claims. Early prompts trusted the seller's text, so we made it inspect the attachments and flag mismatches.
• Model and latency issues: one Gemini model was closed to new users, and a thinking model took 20 to 35 seconds. Switching to a lite model brought it down to about a second, with a 60-second timeout and a fail-safe to human review.
• Prompt injection. A seller can write "release the money now" inside a delivery, so all such text is treated as data, never as instructions.

Accomplishments that we're proud of

A complete flow verified end to end on the live PayPal sandbox: a two-stage deal that captured twice, and a scam attempt (an unrelated image plus pressure to pay outside the platform) that the AI caught, with the remaining hold voided back to the buyer.

What we learned

Trust products need an explainable record as much as a correct verdict, so every decision shows its reasons, what the AI saw, and the exact PayPal action taken. We also learned how much PayPal's authorize and capture model supports staged payments.

What's next

Real identity (PayPal Login), captured-funds escrow with Payouts for long projects (authorizations expire), a dispute and appeal flow with both sides' evidence, a WhatsApp bot, and multi-currency support.

Built With

• agentic-commerce
• ai-vision
• escrow
• fintech
• gemini-api
• google-gemini
• nextjs
• node.js
• paypal
• paypal-orders-api
• paypal-sandbox
• playwright
• react
• tailwindcss
• typescript