Why they stopped
i proved enough of the voice workflow to justify building the real product, then the bottleneck changed. it stopped being a model or latency problem and became a distribution, merchant onboarding, and production-ops problem across shopify, calling, and analytics, and i did not want to keep carrying that full stack by myself.
What it needs
this needs someone who already knows how to sell into shopify brands, or a merchant operator willing to dogfood it on their own store first. the clearest wedge is post-delivery feedback for physical-product merchants, with the checkout-feedback flow used as a second workflow after one merchant segment is working repeatedly.
Hard to replicate
the hard part already done is the cross-system wiring: shopify auth, delivery-triggered call tasks, outbound voice flows, structured transcript + score capture, convex storage, and a real merchant-facing frontend instead of a bare backend demo.
About the project
nina is a shopify feedback product with two surfaces: post-delivery calls for merchants and a checkout-feedback flow for buyers. the bundle includes a next.js frontend, a fastapi backend, convex data storage, shopify oauth and webhook wiring, call task orchestration, transcript + score capture, analytics views, and a checkout extension experiment. there is also an earlier twilio/openai mvp test in the root folder that shows where the product started before it expanded into the fuller merchant workflow.