Inspiration

What it does

How we built it

Inspiration

Freelancers get ghosted after delivering. Clients pay up front and get nothing. Escrow exists, but it still depends on a human to decide whether the work is done, which is slow, biased and expensive. We asked: what if the escrow could judge the work itself?

What it does

AutoSovereign is milestone-based escrow where PayPal holds the money and an AI agent decides when it is earned.

  1. A contract is split into milestones, each with acceptance criteria.
  2. The buyer funds the full amount through PayPal. The contract moves to LOCKED_IN_ESCROW.
  3. The freelancer submits evidence for each milestone.
  4. An AI agent checks the evidence against the criteria and returns approve or reject, with reasoning.
  5. Once every milestone is verified, the payout is released to the freelancer through PayPal Payouts.

Vague evidence is rejected, and the AI explains what is missing. A payout requested before verification is blocked on the server with HTTP 422, so the rule can't be bypassed from the UI.

How we built it

  • PayPal (sandbox): Checkout/Orders funds the escrow, and the capture writes the funded state back to the database. The Payouts API releases the money and returns a real batch ID.
  • AI: a verification agent behind /api/agent/ai-verify reads the milestone criteria and the submitted evidence and produces a verdict. That verdict is what unlocks the payout.
  • Backend: API routes for contracts, funding, verification and payout, backed by a seeded database.

Challenges we ran into

  • Keeping contract state consistent between the PayPal capture and our own database.
  • Making the AI strict enough to reject vague evidence, but fair enough to approve genuine work.
  • Enforcing the payout rule on the server so the AI verdict, not the client, controls the money.

Accomplishments that we're proud of

A complete working loop: real sandbox funding, an AI rejection followed by an approval, a server-enforced guardrail, and a real payout that returns a PayPal batch ID.

What we learned

When an AI decision sits directly in front of a money flow, the guardrails have to live on the server. We also learned that clear, specific acceptance criteria make AI verification much more reliable.

What's next

  • Human dispute escalation when either side contests a verdict
  • Multi-currency payouts and partial releases per milestone
  • Agent-to-agent contracts for agentic commerce ## Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for AutoSovereign

Built With

Share this project:

Updates

Submission history