RelayBuy
Evidence-bound approvals for AI purchases
Created on 1st August 2026
•
RelayBuy
Evidence-bound approvals for AI purchases
The problem RelayBuy solves
AI agents can understand a purchase request and still choose the wrong variant, stale quote, or changed total.
RelayBuy checks OpenAI-extracted intent against live merchant data and version-bound Senso policy, then uses deterministic rules and one human approval before Prava is reachable.
Today it gives teams a clear refusal or approval trail for constrained purchases without handing spending authority to an LLM.
Next I will add authenticated organizations, approver roles, more merchants, and a production Prava flow after the complete sandbox lifecycle is verified.
Challenges I ran into
Three integration bugs shaped RelayBuy. Node fetch intermittently timed out while curl could reach Prava, so I added bounded read-only probes and fail-closed readiness instead of retrying payment calls. One create-session request returned no HTTP status or response ID; RelayBuy persisted an UNKNOWN outcome, disabled duplicate retry, and disarmed payment. Senso records expire quickly, so authorization is bound to content ID, version ID, digest, and freshness, with an atomic refresh-and-verify command. I also isolated private operator flags from Playwright and added connected tests for both the wrong-candidate refusal and exact-candidate approval paths.
Tracks Applied (5)
Best Visa Intelligent Commerce Implementation
Visa
Most Startup-Ready Product
Localhost
OpenAI
OpenAI
Agent Commerce Discovery & Trust
Senso
Best Agentic User Experience
Cheer Project
Cheering for a project means supporting a project you like with as little as 0.0025 ETH. Right now, you can Cheer using ETH on Arbitrum, Optimism and Base.

