Inspiration
Data centers are the physical bottleneck of the AI boom, and site selection is slow, manual, and capital-intensive. We asked: what happens when AI agents do not just recommend where to build, but can act on their own verdicts, including putting money down? And what does it take to trust an agent with a wallet?
What it does
AI Infrastructure SimCity scouts and scores data center sites (power, water, land, connectivity, community sentiment), renders Go / Conditional / No-go verdicts, and then, new for this hackathon, pays for its decisions. An acquisition agent converts a verdict into a land-option deposit through PayPal: full option deposits for Go sites, smaller held deposits for Conditional sites, and nothing, ever, for a No-go. Every payment passes a spending policy first: a $1M portfolio budget, a $250K per-transaction cap, auto-approval under $25K, and human approval above it. Deposits have a full lifecycle: when a Conditional site's conditions clear, its held deposit releases and tops up to the full option; when a site flips to No-go, the deposit is refunded through PayPal. Every movement lands in an auditable deals ledger built on AG Grid, next to a policy console that shows the budget, live exposure, and every policy decision.
How we built it
Next.js and TypeScript, Prisma over Postgres, deployed on Vercel as a separate project for this hackathon. The PayPal integration uses the Orders API v2 (sandbox) with server-side create and capture, plus capture refunds for the lifecycle. The spending policy is a deterministic gate evaluated before any order exists; every decision (auto-approved, human approval, denied) is recorded on the payment and in the agent run log. The ledger UI is AG Grid's React data grid.
Challenges we ran into
Designing the policy gate so an agent can move fast without being able to move wrong: idempotency (one live deposit per assessment), partial states (a held deposit is neither a full option nor a refund), and making refunds and releases first-class flows rather than edge cases.
Accomplishments that we're proud of
An agent with a wallet and a leash: it can commit real (sandbox) money to its own analysis in seconds, but it cannot fund a No-go site, exceed its caps, or spend without leaving an audit trail. Agents propose, policy constrains, humans approve, and the ledger remembers.
What we learned
Agentic commerce is less about the payment call and more about the controls around it: budgets, thresholds, lifecycle states, and reversibility are what make an autonomous payment trustworthy.
What's next
Real-money pilots with stricter approval chains, multi-agent negotiation with landowners, and policy templates for different risk appetites.
Built With
- nextjs
- openai-compatible
- paypal
- postgresql
- prisma
- tailwindcss
- typescript
- vercel
